![NCCloud: Applying Network Coding for the Storage Repair in a
Cloud-of-Clouds
Yuchong Hu†
, Henry C. H. Chen†
, Patrick P. C. Lee†
, Yang Tang‡
†
The Chinese University of Hong Kong, ‡
Columbia University
ychu@inc.cuhk.edu.hk, {chchen,pclee}@cse.cuhk.edu.hk, ty@cs.columbia.edu
Abstract
To provide fault tolerance for cloud storage, recent stud-
ies propose to stripe data across multiple cloud vendors.
However, if a cloud suffers from a permanent failure and
loses all its data, then we need to repair the lost data from
other surviving clouds to preserve data redundancy. We
present a proxy-based system for multiple-cloud storage
called NCCloud, which aims to achieve cost-effective re-
pair for a permanent single-cloud failure. NCCloud is
built on top of network-coding-based storage schemes
called regenerating codes. Specifically, we propose
an implementable design for the functional minimum-
storage regenerating code (F-MSR), which maintains the
same data redundancy level and same storage require-
ment as in traditional erasure codes (e.g., RAID-6), but
uses less repair traffic. We implement a proof-of-concept
prototype of NCCloud and deploy it atop local and com-
mercial clouds. We validate the cost effectiveness of F-
MSR in storage repair over RAID-6, and show that both
schemes have comparable response time performance in
normal cloud storage operations.
1 Introduction
Cloud storage provides an on-demand remote backup so-
lution. However, using a single cloud storage vendor
raises concerns such as having a single point of failure
[3] and vendor lock-ins [1]. As suggested in [1, 3], a
plausible solution is to stripe data across different cloud
vendors. While striping data with conventional erasure
codes performs well when some clouds experience short-
term failures or foreseeable permanent failures [1], there
are real-life cases showing that permanent failures do oc-
cur and are not always foreseeable [23, 14, 20].
This work focuses on unexpected cloud failures.
When a cloud fails permanently, it is important to ac-
tivate storage repair to maintain the level of data re-
dundancy. A repair operation reads data from existing
surviving clouds and reconstructs the lost data in a new
cloud. It is desirable to reduce the repair traffic, and
hence the monetary cost, due to data migration.
Recent studies (e.g., [6, 8, 16, 25]) propose regener-
ating codes for distributed storage. Regenerating codes
are built on the concept of network coding [2]. They aim
to intelligently mix data blocks that are stored in existing
storage nodes, and then regenerate data at a new storage
node. It is shown that regenerating codes reduce the data
repair traffic over traditional erasure codes subject to the
same fault-tolerance level. Despite the favorable prop-
erty, regenerating codes are mainly studied in the theo-
retical context. It remains uncertain regarding the prac-
tical performance of regenerating codes, especially with
the encoding overhead incurred in regenerating codes.
In this paper, we propose NCCloud, a proxy-based
system designed for multiple-cloud storage. We pro-
pose the first implementable design for the functional
minimum-storage regenerating code (F-MSR) [8], and in
particular, we eliminate the need of performing encoding
operations within storage nodes as in existing theoretical
studies. Our F-MSR implementation maintains double-
fault tolerance and has the same storage cost as in tra-
ditional erasure coding schemes based on RAID-6, but
uses less repair traffic when recovering a single-cloud
failure. On the other hand, unlike most erasure coding
schemes that are systematic (i.e., original data chunks are
kept), F-MSR is non-systematic and stores only linearly
combined code chunks. Nevertheless, F-MSR is suited
to rarely-read long-term archival applications [6].
We show that in a practical deployment setting, F-
MSR can save the repair cost by 25% compared to
RAID-6 for a four-cloud setting, and up to 50% as the
number of clouds further increases. In addition, we con-
duct extensive evaluations on both local cloud and com-
mercial cloud settings. We show that our F-MSR imple-
mentation only adds a small encoding overhead, which
can be easily masked by the file transfer time over the
Internet. Thus, our work validates the practicality of F-
MSR via NCCloud, and motivates further studies of re-
alizing regenerating codes in large-scale deployments.
2 Motivation of F-MSR
We consider a distributed, multiple-cloud storage setting
from a client’s perspective, such that we stripe data over
multiple cloud vendors. We propose a proxy-based de-
sign [1] that interconnects multiple cloud repositories, as
shown in Figure 1(a). The proxy serves as an interface
between client applications and the clouds. If a cloud
experiences a permanent failure, the proxy activates the
repair operation, as shown in Figure 1(b). That is, the](https://image.slidesharecdn.com/ieeeprotechnosolutions-ieeejavaproject-nccloudapplyingnetworkcodingforthestoragerepairina-160127124733/85/Ieeepro-techno-solutions-ieee-java-project-nc-cloud-applying-network-coding-for-the-storage-repair-in-a-1-320.jpg)
![Cloud 2
Cloud 1
Cloud 3
Cloud 4
Cloud 2
Cloud 1
Cloud 3
Cloud 4
(a) Normal operation (b) Repair operation
Cloud 5
Proxy Proxy
Figure 1: Proxy-based design for multiple-cloud stor-
age: (a) normal operation, and (b) repair operation when
Cloud node 1 fails. During repair, the proxy regenerates
data for the new cloud.
proxy reads the essential data pieces from other surviv-
ing clouds, reconstructs new data pieces, and writes these
new pieces to a new cloud. Note that this repair operation
does not involve direct interactions among the clouds.
We consider fault-tolerant storage based on maximum
distance separable (MDS) codes. Given a file object, we
divide it into equal-size native chunks, which in a non-
coded system, would be stored on k clouds. With cod-
ing, the native chunks are encoded by linear combina-
tions to form code chunks. The native and code chunks
are distributed over n > k clouds. When an MDS code is
used, the original file object may be reconstructed from
the chunks contained in any k of the n clouds. Thus, it
tolerates the failure of any n − k clouds. We call this
feature the MDS property. The extra feature of F-MSR
is that reconstructing a single native or code chunk may
be achieved by reading up to 50% less data from the sur-
viving clouds than reconstructing the whole file.
This paper considers a multiple-cloud setting that is
double-fault tolerant (e.g., RAID-6) and provides data
availability toward at most two cloud failures (e.g., a few
days of outages [7]). That is, we set k = n − 2. We
expect that such a fault tolerance level suffices in prac-
tice. Given that a permanent failure is less frequent but
possible, our primary objective is to minimize the cost of
storage repair for a permanent single-cloud failure, due
to the migration of data over the clouds.
We define the repair traffic as the amount of outbound
data being read from other surviving clouds during the
single-cloud failure recovery. Our goal is to minimize
the repair traffic for cost-effective repair. Here, we do not
consider the inbound traffic (i.e., the data being written
to a cloud), as it is free of charge in many cloud vendors
(see Table 1 in Section 5).
We now show how F-MSR saves the repair traffic via
an example. Suppose that we store a file of size M on
four clouds, each viewed as a logical storage node. Let
us first consider RAID-6, which is double-fault tolerant.
Here, we consider the RAID-6 implementation based on
Reed-Solomon codes [26], as shown in Figure 2(a). We
divide the file into two native chunks (i.e., A and B) of
size M/2 each. We add two code chunks formed by the
linear combinations of the native chunks. Suppose now
that Node 1 is down. Then the proxy must download
the same number of chunks as the original file from two
other nodes (e.g., B and A + B from Nodes 2 and 3, re-
spectively). It then reconstructs and stores the lost chunk
A on the new node. The total storage size is 2M, while
the repair traffic is M.
We now consider the double-fault tolerant implemen-
tation of F-MSR in a proxy-based setting, as shown
in Figure 2(b). F-MSR divides the file into four na-
tive chunks, and constructs eight distinct code chunks
P1, · · · , P8 formed by different linear combinations of
the native chunks. Each code chunk has the same size
M/4 as a native chunk. Any two nodes can be used to re-
cover the original four native chunks. Suppose Node 1 is
down. The proxy collects one code chunk from each sur-
viving node, so it downloads three code chunks of size
M/4 each. Then the proxy regenerates two code chunks
P1 and P2 formed by different linear combinations of
the three code chunks. Note that P1 and P2 are still lin-
ear combinations of the native chunks. The proxy then
writes P1 and P2 to the new node. In F-MSR, the stor-
age size is 2M (as in RAID-6), but the repair traffic is
0.75M, which is 25% of saving.
To generalize F-MSR for n storage nodes, we divide a
file of size M into 2(n − 2) native chunks, and generate
4(n − 2) code chunks. Then each node will store two
code chunks of size M
2(n−2) each. Thus, the total storage
size is Mn
n−2 . To repair a failed node, we download one
chunk from each of n − 1 nodes, so the repair traffic is
M(n−1)
2(n−2) . In contrast, for RAID-6, the total storage size is
also Mn
n−2 , while the repair traffic is M. When n is large,
F-MSR can save the repair traffic by close to 50%.
Note that F-MSR keeps only code chunks rather than
native chunks. To access a single chunk of a file, we need
to download and decode the entire file for the particular
chunk. Nevertheless, F-MSR is acceptable for long-term
archival applications, whose read frequency is typically
low [6]. Also, to restore backups, it is natural to retrieve
the entire file rather than a particular chunk.
This paper considers the baseline RAID-6 implemen-
tation using Reed-Solomon codes. Its repair method is to
reconstruct the whole file, and is applicable for all era-
sure codes in general. Recent studies [18, 28, 29] show
that data reads can be minimized specifically for XOR-
based erasure codes. For example, in RAID-6, data reads
can be reduced by 25% compared to reconstructing the
whole file [28, 29]. Although such approaches are sub-
optimal (recall that F-MSR can save up to 50% of repair
traffic in RAID-6), their use of efficient XOR operations
can be of practical interest.](https://image.slidesharecdn.com/ieeeprotechnosolutions-ieeejavaproject-nccloudapplyingnetworkcodingforthestoragerepairina-160127124733/85/Ieeepro-techno-solutions-ieee-java-project-nc-cloud-applying-network-coding-for-the-storage-repair-in-a-2-320.jpg)
![ANode 1
Proxy
BNode 2
A+BNode 3
A+2BNode 4
A
B
Object of size M
B
A+B
A
New node
A
Node 1
Proxy
Node 2
Node 3
Node 4
Object of size M
New node
A
B
C
D
P1
P2
P3
P4
P7
P8
P5
P6
P3
P5
P7
P1’
P2’
P1’
P2’
(a) RAID-6 (b) F-MSR
Figure 2: Examples of repair operations in RAID-6 and F-MSR with n = 4 and k = 2.
3 F-MSR Implementation
In this section, we present a systematic approach for im-
plementing F-MSR. We specify three operations for F-
MSR on a particular file object: (1) file upload; (2) file
download; (3) repair. A key difference of our imple-
mentation from prior theoretical studies is that we do
not require storage nodes to have encoding capabilities,
so our implementation can be compatible with today’s
cloud storage. Another key design issue is that instead of
simply generating random linear combinations for code
chunks (as assumed in [8]), we also guarantee that the
generated linear combinations always satisfy the MDS
property to ensure data availability, even after iterative
repairs. Here, we implement F-MSR as an MDS code
for general (n,k). We assume that each cloud repository
corresponds to a logical storage node.
3.1 File Upload
To upload a file F, we first divide it into k(n − k) equal-
size native chunks, denoted by (Fi)i=1,2,···,k(n−k). We
then encode these k(n − k) native chunks into n(n − k)
code chunks, denoted by (Pi)i=1,2,···,n(n−k). Each Pi
is formed by a linear combination of the k(n − k) na-
tive chunks. Specifically, we let EM = [αi,j] be an
n(n−k)×k(n−k) encoding matrix for some coefficients
αij (where i = 1, . . . , n(n−k) and j = 1, . . . , k(n−k))
in the Galois field GF(28
). We call a row vector of
EM an encoding coefficient vector (ECV), which con-
tains k(n − k) elements. We let ECVi denote the ith
row vector of EM. We compute each Pi by the scalar
product of ECVi and the native chunk vector (Fi), i.e.,
Pi =
k(n−k)
j=1 αijFj for i = 1, 2, · · · , n(n − k), where
all arithmetic operations are performed over GF(28
). The
code chunks are then evenly stored in the n storage
nodes, each having (n − k) chunks. Also, we store the
whole EM in a metadata object that is then replicated to
all storage nodes (see Section 4). There are many ways
of constructing EM, as long as it satisfies the MDS prop-
erty and the repair MDS property (see Section 3.3). Note
that the implementation details of the arithmetic opera-
tions in Galois Fields are extensively discussed in [15].
3.2 File Download
To download a file, we first download the correspond-
ing metadata object that contains the ECVs. Then we
select any k of the n storage nodes, and download the
k(n − k) code chunks from the k nodes. The ECVs of
the k(n−k) code chunks can form a k(n−k)×k(n−k)
square matrix. If the MDS property is maintained, then
by definition, the inverse of the square matrix must ex-
ist. Thus, we multiply the inverse of the square matrix
with the code chunks and obtain the original k(n − k)
native chunks. The idea is that we treat F-MSR as a stan-
dard Reed-Solomon code, and our technique of creating
an inverse matrix to decode the original data has been
described in the tutorial [22].
3.3 Iterative Repairs
We now consider the repair of F-MSR for a file F for a
permanent single-node failure. Given that F-MSR regen-
erates different chunks in each repair, one challenge is to
ensure that the MDS property still holds even after itera-
tive repairs. This is in contrast to regenerating the exact
lost chunks as in RAID-6, which guarantees the invari-
ance of the stored chunks. Here, we propose a two-phase
checking heuristic as follows. Suppose that the (r − 1)th
repair is successful, and we now consider how to operate
the rth
repair for a single permanent node failure (where
r ≥ 1). We first check if the new set of chunks in all stor-
age nodes satisfies the MDS property after the rth
repair.
In addition, we also check if another new set of chunks in
all storage nodes still satisfies the MDS property after the
(r + 1)th
repair, should another single permanent node
failure occur (we call this the repair MDS property). We
now describe the rth
repair as follows.
Step 1: Download the encoding matrix from a surviving
node. Recall that the encoding matrix EM specifies the
ECVs for constructing all code chunks via linear combi-
nations of native chunks. We use these ECVs for our later
two-phase checking heuristic. Since we embed EM in a
metadata object that is replicated, we can simply down-
load the metadata object from one of the surviving nodes.
Step 2: Select one random ECV from each of the n − 1
surviving nodes. Note that each ECV in EM corre-](https://image.slidesharecdn.com/ieeeprotechnosolutions-ieeejavaproject-nccloudapplyingnetworkcodingforthestoragerepairina-160127124733/85/Ieeepro-techno-solutions-ieee-java-project-nc-cloud-applying-network-coding-for-the-storage-repair-in-a-3-320.jpg)
![sponds to a code chunk. We randomly pick one ECV
from each of the n − 1 surviving nodes. We call these
ECVs to be ECVi1
, ECVi2
, · · ·, ECVin−1
.
Step 3: Generate a repair matrix. We construct a (n −
k)×(n−1) repair matrix RM = [γi,j], where each ele-
ment γi,j (where i = 1, . . . , n−k and j = 1, . . . , n−1) is
randomly selected in GF(28
). Note that the idea of gen-
erating a random matrix for reliable storage is consistent
with that in [24].
Step 4: Compute the ECVs for the new code chunks and
reproduce a new encoding matrix. We multiply RM
with the ECVs selected in Step 2 to construct n − k
new ECVs, denoted by ECVi =
n−1
j=1 γi,jECVij
for
i = 1, 2, · · · , n − k. Then we reproduce a new encoding
matrix, denoted by EM , which is given by:
ith
row vector of EM =
ECVi, i is a surviving node,
ECVi, i is a new node.
Step 5: Given EM , check if both the MDS and repair
MDS properties are satisfied. Intuitively, we verify the
MDS property by enumerating all n
k subsets of k nodes
to see if each of their corresponding encoding matrices
forms a full rank. For the repair MDS property, we check
that for any failed node (out of n nodes), we can collect
any one out of n−k chunks from the other n−1 surviving
nodes and reconstruct the chunks in the new node, such
that the MDS property is maintained. The number of
checks performed for the repair MDS property is at most
n(n − k)n−1 n
k . If n is small, then the enumeration
complexities for both MDS and repair MDS properties
are manageable. If either one phase fails, then we return
to Step 2 and repeat. We emphasize that Steps 1 to 5 only
deal with the ECVs, so their overhead does not depend
on the chunk size.
Step 6: Download the actual chunk data and regenerate
new chunk data. If the two-phase checking in Step 5
succeeds, then we proceed to download the n−1 chunks
that correspond to the selected ECVs in Step 2 from the
n − 1 surviving storage nodes to NCCloud. Also, using
the new ECVs computed in Step 4, we regenerate new
chunks and upload them from NCCloud to a new node.
Remark: We claim that in addition to checking the MDS
property, checking the repair MDS property is essential
for iterative repairs. We conduct simulations to justify
that checking the repair MDS property can make itera-
tive repairs sustainable. In our simulations, we consider
multiple rounds of permanent node failures for different
values of n. Specifically, in each round, we randomly
pick a node to permanently fail and trigger a repair. We
say a repair is bad if the loop of Steps 2 to 5 is repeated
over 10 times. We observe that without checking the re-
pair MDS property, we see a bad repair very quickly,
say after no more than 7 and 2 rounds for n = 8 and n
= 12, respectively. On the other hand, checking the re-
pair MDS property makes iterative repairs sustainable for
hundreds of rounds for different values of n, and we do
not yet find any bad repair after extensive simulations.
4 NCCloud Design and Implementation
We implement NCCloud as a proxy that bridges user
applications and multiple clouds. Its design is built on
three layers. The file system layer presents NCCloud as a
mounted drive, which can thus be easily interfaced with
general user applications. The coding layer deals with
the encoding and decoding functions. The storage layer
deals with read/write requests with different clouds.
Each file is associated with a metadata object, which is
replicated at each repository. The metadata object holds
the file details and the coding information (e.g., encoding
coefficients for F-MSR).
NCCloud is mainly implemented in Python, while the
storage schemes are implemented in C for better effi-
ciency. The file system layer is built on FUSE [12].
The coding layer implements both RAID-6 and F-MSR.
RAID-6 is built on zfec [30], and our F-MSR implemen-
tation mimics the optimizations made in zfec for a fair
comparison.
Recall that F-MSR generates multiple chunks to be
stored on the same repository. To save the request cost
overhead (see Table 1), multiple chunks destined for the
same repository are aggregated before upload. Thus,
F-MSR keeps only one (aggregated) chunk per file ob-
ject on each cloud, as in RAID-6. To retrieve a specific
chunk, we calculate its offset within the combined chunk
and issue a range GET request.
5 Evaluation
We now use our NCCloud prototype to evaluate RAID-6
(based on Reed-Solomon codes) and F-MSR in multiple-
cloud storage. In particular, we focus on the setting n =
4 and k = 2. We expect that using n = 4 clouds may
suffice for practical deployment. Based on this setting,
we allow data retrieval with at most two cloud failures.
The goal of our experiments is to explore the practi-
cality of using F-MSR in multiple-cloud storage. Our
evaluation consists of two parts. We first compare the
monetary costs of using RAID-6 and F-MSR based on
the price plans of today’s cloud vendors. We also em-
pirically evaluate the response time performance of our
NCCloud prototype atop a local cloud and also a com-
mercial cloud vendor.
5.1 Cost Analysis
Table 1 shows the monthly price plans for three major
vendors as of September 2011. For Amazon S3, we take
the cost from the first chargeable usage tier (i.e., storage](https://image.slidesharecdn.com/ieeeprotechnosolutions-ieeejavaproject-nccloudapplyingnetworkcodingforthestoragerepairina-160127124733/85/Ieeepro-techno-solutions-ieee-java-project-nc-cloud-applying-network-coding-for-the-storage-repair-in-a-4-320.jpg)
![S3 RS Azure
Storage (per GB) $0.14 $0.15 $0.15
Data transfer in (per GB) free free free
Data transfer out (per GB) $0.12 $0.18 $0.15
PUT,POST (per 10K requests) $0.10 free $0.01
GET (per 10K requests) $0.01 free $0.01
Table 1: Monthly price plans (in US dollars) for Amazon
S3 (US Standard), Rackspace Cloud Files and Windows
Azure Storage, as of September, 2011.
usage within 1TB/month; data transferred out more than
1GB/month but less than 10TB/month).
From the analysis in Section 2, we can save 25% of
the download traffic during storage repair when n = 4.
The storage size and the number of chunks being gen-
erated per file object are the same in both RAID-6 and
F-MSR (assuming that we aggregate chunks in F-MSR
as described in Section 4). However, in the analysis,
we have ignored two practical considerations: the size
of metadata (Section 4) and the number of requests is-
sued during repair. We now argue that they are negligi-
ble and that the simplified calculations based only on file
size suffice for real-life applications.
Metadata size: Our implementation currently keeps the
F-MSR metadata size within 160B, regardless of the file
size. NCCloud aims at long-term backups (see Sec-
tion 2), and can be integrated with other backup applica-
tions. Existing backup applications (e.g., [27, 11]) typi-
cally aggregate small files into a larger data chunk in or-
der to save the processing overhead. For example, the de-
fault setting for Cumulus [27] creates chunks of around
4MB each. The metadata size is thus usually negligible.
Number of requests: From Table 1, we see that some
cloud vendors nowadays charge for requests. RAID-6
and F-MSR differ in the number of requests when re-
trieving data during repair. Suppose that we store a file
of size 4MB with n = 4 and k = 2. During repair,
RAID-6 and F-MSR retrieve two and three chunks, re-
spectively (see Figure 2). The cost overhead due to the
GET request for RAID-6 is at most 0.427%, and that for
F-MSR is at most 0.854%, a mere 0.427% increase.
5.2 Response Time Analysis
We deploy our NCCloud prototype in real environments.
We then evaluate the response time performance of dif-
ferent operations in two scenarios. The first part ana-
lyzes in detail the time taken by different NCCloud op-
erations, and is done on a local cloud storage in order
to lessen the effects of network fluctuations. The second
part evaluates how NCCloud actually performs in com-
mercial clouds. All results are averaged over 40 runs.
0
10
20
30
40
50
500 400 300 200 100 50 10 1
Responsetime
(seconds)
File size (MB)
RAID-6
F-MSR
(a) File upload
0
2
4
6
8
10
12
500 400 300 200 100 50 10 1
Responsetime
(seconds)
File size (MB)
RAID-6
F-MSR
(b) File download
0
5
10
15
20
25
30
35
500 400 300 200 100 50 10 1
Responsetime
(seconds)
File size (MB)
RAID-6 (code chunk repair)
RAID-6 (native chunk repair)
F-MSR
(c) Repair
Figure 3: Response times of main NCCloud operations.
0
10
20
30
40
50
60
RAID-6
F-MSR
RAID-6
F-MSR
R6(code)
R6(native)
F-MSR
Timetaken(seconds)
(a) Upload (b) Download (c) Repair
File I/O
Upload
Download
Encoding
Decoding
Figure 4: Breakdown of response time when dealing with
500MB file.
5.2.1 On a Local Cloud
The experiment on a local cloud is carried out on an
object-based storage platform based on OpenStack Swift
1.4.2 [21]. NCCloud is mounted on a machine with Intel
Xeon E5620 and 16GB RAM. This machine is connected
to an OpenStack Swift platform attached with a number
of storage servers, each with Intel Core i5-2400 and 8GB
RAM. We create (n+1) = 5 containers on Swift, so each
container resembles a cloud repository (one of them is a
spare node used in repair).
In this experiment, we test the response time of three
basic operations of NCCloud: (a) file upload; (b) file
download; (c) repair. We use eight randomly generated
files from 1MB to 500MB as the data set. We set the path
of a chosen repository to a non-existing location to simu-
late a node failure in repair. Note that there are two types
of repair for RAID-6, depending on whether the failed
node contains a native chunk or a code chunk.](https://image.slidesharecdn.com/ieeeprotechnosolutions-ieeejavaproject-nccloudapplyingnetworkcodingforthestoragerepairina-160127124733/85/Ieeepro-techno-solutions-ieee-java-project-nc-cloud-applying-network-coding-for-the-storage-repair-in-a-5-320.jpg)
![0
1
2
3
4
5
6
10 5 2 1
Responsetime
(seconds)
File size (MB)
RAID-6
F-MSR
(a) File upload
0
0.5
1
1.5
2
2.5
3
10 5 2 1
Responsetime
(seconds)
File size (MB)
RAID-6
F-MSR
(b) File download
0
1
2
3
4
5
6
7
10 5 2 1
Responsetime
(seconds)
File size (MB)
RAID-6 (code chunk repair)
RAID-6 (native chunk repair)
F-MSR
(c) Repair
Figure 5: Response times of NCCloud on Azure.
Figure 3 shows the response times of all three op-
erations (with 95% confidence intervals plotted), and
Figure 4 shows five key constituents of the response
time when dealing with a 500MB file. Figure 3 shows
that RAID-6 has less response time in file upload and
download. With the help of Figure 4, we pinpoint the
overhead of F-MSR over RAID-6. Due to having the
same MDS property, RAID-6 and F-MSR exhibit similar
data transfer time during upload/download. However, F-
MSR displays a noticeable encoding/decoding overhead
over RAID-6. When uploading a 500MB file, RAID-6
takes 1.490s to encode while F-MSR takes 5.365s; when
downloading a 500MB file, no decoding is needed in the
case of RAID-6 as the native chunks are available, but
F-MSR takes 2.594s to decode.
On the other hand, F-MSR has slightly less response
time in repair. The main advantage of F-MSR is that it
needs to download less data during repair. In repairing
a 500MB file, F-MSR spends 3.887s in download, while
the native-chunk repair of RAID-6 spends 4.832s.
Although RAID-6 generally has less response time
than F-MSR in a local cloud environment, we expect that
the encoding/decoding overhead of F-MSR can be easily
masked by network fluctuations over the Internet, as will
be shown next.
5.2.2 On a Commercial Cloud
The following experiment is carried out on a machine
with Intel Xeon E5530 and 16GB RAM running 64-bit
Ubuntu 9.10. We repeat the three operations in Sec-
tion 5.2.1 on four randomly generated files from 1MB to
10MB atop Windows Azure Storage. On Azure, we cre-
ate (n+1) = 5 containers to mimic different cloud repos-
itories. The same operation for both RAID-6 and F-MSR
are run interleaved to lessen the effect of network fluctu-
ation on the comparison due to different times of the day.
Figure 5 shows the results for different file sizes with
95% confidence intervals plotted. Note that although we
have used only Azure in this experiment, actual usage of
NCCloud should stripe data over different vendors and
locations for better availability guarantees.
From Figure 5, we do not see distinct response time
differences between RAID-6 and F-MSR in all opera-
tions. Furthermore, on the same machine, F-MSR takes
around 0.150s to encode and 0.064s to decode a 10MB
file (not shown in the figures). These constitute roughly
3% of the total upload and download times (4.962s and
2.240s respectively). Given that the 95% confidence in-
tervals for the upload and download times are 0.550s
and 0.438s respectively, network fluctuation plays a big-
ger role in determining the response time. Overall, we
demonstrate that F-MSR does not have significant per-
formance overhead over our baseline RAID-6 implemen-
tation.
6 Related Work
There are several systems proposed for multiple-cloud
storage. HAIL [5] provides integrity and availabil-
ity guarantees for stored data. DEPSKY [4] addresses
Byzantine Fault Tolerance by combining encryption and
erasure coding for stored data. RACS [1] uses erasure
coding to mitigate vendor lock-ins when switching cloud
vendors. It retrieves data from the cloud that is about
to be failed and moves the data to the new cloud. Un-
like RACS, NCCloud excludes the failed cloud in repair.
All the above systems are based on erasure codes, while
NCCloud considers regenerating codes with an emphasis
on storage repair.
Regenerating codes (see survey [9]) exploit the opti-
mal trade-off between storage cost and repair traffic. Ex-
isting studies mainly focus on theoretical analysis. Sev-
eral studies (e.g., [10, 13, 19]) empirically evaluate ran-
dom linear codes for peer-to-peer storage. However,
their evaluations are mainly based on simulations. NCFS
[17] implements regenerating codes, but does not con-
sider MSR codes that are based on linear combinations.
Here, we consider the F-MSR implementation, and per-
form empirical experiments in multiple-cloud storage.
7 Conclusions
We present NCCloud, a multiple-cloud storage file sys-
tem that practically addresses the reliability of today’s
cloud storage. NCCloud not only achieves fault tolerance
of storage, but also allows cost-effective repair when a
cloud permanently fails. NCCloud implements a practi-
cal version of the functional minimum storage regenerat-
ing code (F-MSR), which regenerates new chunks during](https://image.slidesharecdn.com/ieeeprotechnosolutions-ieeejavaproject-nccloudapplyingnetworkcodingforthestoragerepairina-160127124733/85/Ieeepro-techno-solutions-ieee-java-project-nc-cloud-applying-network-coding-for-the-storage-repair-in-a-6-320.jpg)
![repair subject to the required degree of data redundancy.
Our NCCloud prototype shows the effectiveness of F-
MSR in accessing data, in terms of monetary costs and
response times. The source code of NCCloud is available
at http://ansrlab.cse.cuhk.edu.hk/software/nccloud.
8 Acknowledgments
We would like to thank our shepherd, James Plank, and
the anonymous reviewers for their valuable comments.
This work was supported by grant AoE/E-02/08 from the
University Grants Committee of Hong Kong.
References
[1] H. Abu-Libdeh, L. Princehouse, and H. Weather-
spoon. RACS: A Case for Cloud Storage Diversity.
In Proc. of ACM SoCC, 2010.
[2] R. Ahlswede, N. Cai, S.-Y. R. Li, and R. W. Ye-
ung. Network Information Flow. IEEE Trans. on
Information Theory, 46(4):1204–1216, Jul 2000.
[3] M. Armbrust, A. Fox, R. Griffith, A. D. Joseph,
R. Katz, A. Konwinski, G. Lee, D. Patterson,
A. Rabkin, I. Stoica, and M. Zaharia. A View of
Cloud Computing. Communications of the ACM,
53(4):50–58, 2010.
[4] A. Bessani, M. Correia, B. Quaresma, F. Andr´e,
and P. Sousa. DEPSKY: Dependable and Secure
Storage in a Cloud-of-Clouds. In Proc. of ACM Eu-
roSys, 2011.
[5] K. D. Bowers, A. Juels, and A. Oprea. HAIL:
A High-Availability and Integrity Layer for Cloud
Storage. In Proc. of ACM CCS, 2009.
[6] B. Chen, R. Curtmola, G. Ateniese, and R. Burns.
Remote Data Checking for Network Coding-Based
Distributed Storage Systems. In Proc. of ACM
CCSW, 2010.
[7] CNNMoney. Amazon’s cloud is back, but still hazy.
http://money.cnn.com/2011/04/25/
technology/amazon_cloud/index.htm.
[8] A. G. Dimakis, P. B. Godfrey, Y. Wu, M. Wain-
wright, and K. Ramchandran. Network Coding for
Distributed Storage Systems. IEEE Trans. on In-
formation Theory, 56(9):4539–4551, Sep 2010.
[9] A. G. Dimakis, K. Ramchandran, Y. Wu, and
C. Suh. A Survey on Network Codes for Dis-
tributed Storage. Proc. of the IEEE, 99(3):476–489,
Mar 2011.
[10] A. Duminuco and E. Biersack. A Practical Study of
Regenerating Codes for Peer-to-Peer Backup Sys-
tems. In Proc. of IEEE ICDCS, 2009.
[11] B. Escoto and K. Loafman. Duplicity. http://
duplicity.nongnu.org/.
[12] FUSE. http://fuse.sourceforge.net/.
[13] C. Gkantsidis and P. Rodriguez. Network coding
for large scale content distribution. In Proc. of IN-
FOCOM, 2005.
[14] GmailBlog. Gmail back soon for everyone. http:
//gmailblog.blogspot.com/2011/02/
gmail-back-soon-for-everyone.html.
[15] K. M. Greenan, E. L. Miller, and T. J. E. Schwarz.
Optimizing Galois Field Arithmetic for Diverse
Processor Architectures and Applications. In Proc.
of IEEE MASCOTS, 2008.
[16] Y. Hu, Y. Xu, X. Wang, C. Zhan, and P. Li. Coop-
erative recovery of distributed storage systems from
multiple losses with network coding. IEEE JSAC,
28(2):268–276, Feb 2010.
[17] Y. Hu, C.-M. Yu, Y.-K. Li, P. P. C. Lee, and J. C. S.
Lui. NCFS: On the Practicality and Extensibility of
a Network-Coding-Based Distributed File System.
In Proc. of NetCod, 2011.
[18] O. Khan, R. Burns, J. Plank, and C. Huang. In
Search of I/O-Optimal Recovery from Disk Fail-
ures. In USENIX HotStorage, 2011.
[19] M. Martal´o, M. Picone, M. Amoretti, G. Ferrari,
and R. Raheli. Randomized Network Coding in
Distributed Storage Systems with Layered Overlay.
In Information Theory and Application Workshop,
2011.
[20] E. Naone. Are We Safeguarding Social Data?
http://www.technologyreview.com/
blog/editors/22924/, Feb 2009.
[21] OpenStack Object Storage. http://www.
openstack.org/projects/storage/.
[22] J. S. Plank. A Tutorial on Reed-Solomon Coding
for Fault-Tolerance in RAID-like Systems. Soft-
ware - Practice & Experience, 27(9):995–1012,
Sep 1997.
[23] C. Preimesberger. Many data centers unpre-
pared for disasters: Industry group, Mar 2011.
http://www.eweek.com/c/a/IT-
Management/Many-Data-Centers-
Unprepared-for-Disasters-Industry-
Group-772367/.
[24] M. O. Rabin. Efficient Dispersal of Information
for Security, Load Balancing, and Fault Tolerance.
Journal of the ACM, 36(2):335–348, Apr 1989.
[25] K. V. Rashmi, N. B. Shah, P. V. Kumar, and
K. Ramchandran. Explicit Construction of Optimal
Exact Regenerating Codes for Distributed Storage.
In Proc. of Allerton Conference, 2009.](https://image.slidesharecdn.com/ieeeprotechnosolutions-ieeejavaproject-nccloudapplyingnetworkcodingforthestoragerepairina-160127124733/85/Ieeepro-techno-solutions-ieee-java-project-nc-cloud-applying-network-coding-for-the-storage-repair-in-a-7-320.jpg)
![[26] I. Reed and G. Solomon. Polynomial Codes over
Certain Finite Fields. Journal of the Society for In-
dustrial and Applied Mathematics, 8(2):300–304,
1960.
[27] M. Vrable, S. Savage, and G. Voelker. Cumu-
lus: Filesystem backup to the cloud. In Proc. of
USENIX FAST, 2009.
[28] Z. Wang, A. Dimakis, and J. Bruck. Rebuilding
for Array Codes in Distributed Storage Systems. In
IEEE GLOBECOM Workshops, 2010.
[29] L. Xiang, Y. Xu, J. Lui, Q. Chang, Y. Pan, and R. Li.
A Hybrid Approach to Failed Disk Recovery Using
RAID-6 Codes: Algorithms and Performance Eval-
uation. ACM Trans. on Storage, 7(3):11, 2011.
[30] zfec. http://pypi.python.org/pypi/
zfec.](https://image.slidesharecdn.com/ieeeprotechnosolutions-ieeejavaproject-nccloudapplyingnetworkcodingforthestoragerepairina-160127124733/85/Ieeepro-techno-solutions-ieee-java-project-nc-cloud-applying-network-coding-for-the-storage-repair-in-a-8-320.jpg)
Ieeepro techno solutions ieee java project - nc cloud applying network coding for the storage repair in a
- 1. NCCloud: Applying Network Coding for the Storage Repair in a Cloud-of-Clouds Yuchong Hu† , Henry C. H. Chen† , Patrick P. C. Lee† , Yang Tang‡ † The Chinese University of Hong Kong, ‡ Columbia University ychu@inc.cuhk.edu.hk, {chchen,pclee}@cse.cuhk.edu.hk, ty@cs.columbia.edu Abstract To provide fault tolerance for cloud storage, recent stud- ies propose to stripe data across multiple cloud vendors. However, if a cloud suffers from a permanent failure and loses all its data, then we need to repair the lost data from other surviving clouds to preserve data redundancy. We present a proxy-based system for multiple-cloud storage called NCCloud, which aims to achieve cost-effective re- pair for a permanent single-cloud failure. NCCloud is built on top of network-coding-based storage schemes called regenerating codes. Specifically, we propose an implementable design for the functional minimum- storage regenerating code (F-MSR), which maintains the same data redundancy level and same storage require- ment as in traditional erasure codes (e.g., RAID-6), but uses less repair traffic. We implement a proof-of-concept prototype of NCCloud and deploy it atop local and com- mercial clouds. We validate the cost effectiveness of F- MSR in storage repair over RAID-6, and show that both schemes have comparable response time performance in normal cloud storage operations. 1 Introduction Cloud storage provides an on-demand remote backup so- lution. However, using a single cloud storage vendor raises concerns such as having a single point of failure [3] and vendor lock-ins [1]. As suggested in [1, 3], a plausible solution is to stripe data across different cloud vendors. While striping data with conventional erasure codes performs well when some clouds experience short- term failures or foreseeable permanent failures [1], there are real-life cases showing that permanent failures do oc- cur and are not always foreseeable [23, 14, 20]. This work focuses on unexpected cloud failures. When a cloud fails permanently, it is important to ac- tivate storage repair to maintain the level of data re- dundancy. A repair operation reads data from existing surviving clouds and reconstructs the lost data in a new cloud. It is desirable to reduce the repair traffic, and hence the monetary cost, due to data migration. Recent studies (e.g., [6, 8, 16, 25]) propose regener- ating codes for distributed storage. Regenerating codes are built on the concept of network coding [2]. They aim to intelligently mix data blocks that are stored in existing storage nodes, and then regenerate data at a new storage node. It is shown that regenerating codes reduce the data repair traffic over traditional erasure codes subject to the same fault-tolerance level. Despite the favorable prop- erty, regenerating codes are mainly studied in the theo- retical context. It remains uncertain regarding the prac- tical performance of regenerating codes, especially with the encoding overhead incurred in regenerating codes. In this paper, we propose NCCloud, a proxy-based system designed for multiple-cloud storage. We pro- pose the first implementable design for the functional minimum-storage regenerating code (F-MSR) [8], and in particular, we eliminate the need of performing encoding operations within storage nodes as in existing theoretical studies. Our F-MSR implementation maintains double- fault tolerance and has the same storage cost as in tra- ditional erasure coding schemes based on RAID-6, but uses less repair traffic when recovering a single-cloud failure. On the other hand, unlike most erasure coding schemes that are systematic (i.e., original data chunks are kept), F-MSR is non-systematic and stores only linearly combined code chunks. Nevertheless, F-MSR is suited to rarely-read long-term archival applications [6]. We show that in a practical deployment setting, F- MSR can save the repair cost by 25% compared to RAID-6 for a four-cloud setting, and up to 50% as the number of clouds further increases. In addition, we con- duct extensive evaluations on both local cloud and com- mercial cloud settings. We show that our F-MSR imple- mentation only adds a small encoding overhead, which can be easily masked by the file transfer time over the Internet. Thus, our work validates the practicality of F- MSR via NCCloud, and motivates further studies of re- alizing regenerating codes in large-scale deployments. 2 Motivation of F-MSR We consider a distributed, multiple-cloud storage setting from a client’s perspective, such that we stripe data over multiple cloud vendors. We propose a proxy-based de- sign [1] that interconnects multiple cloud repositories, as shown in Figure 1(a). The proxy serves as an interface between client applications and the clouds. If a cloud experiences a permanent failure, the proxy activates the repair operation, as shown in Figure 1(b). That is, the
- 2. Cloud 2 Cloud 1 Cloud 3 Cloud 4 Cloud 2 Cloud 1 Cloud 3 Cloud 4 (a) Normal operation (b) Repair operation Cloud 5 Proxy Proxy Figure 1: Proxy-based design for multiple-cloud stor- age: (a) normal operation, and (b) repair operation when Cloud node 1 fails. During repair, the proxy regenerates data for the new cloud. proxy reads the essential data pieces from other surviv- ing clouds, reconstructs new data pieces, and writes these new pieces to a new cloud. Note that this repair operation does not involve direct interactions among the clouds. We consider fault-tolerant storage based on maximum distance separable (MDS) codes. Given a file object, we divide it into equal-size native chunks, which in a non- coded system, would be stored on k clouds. With cod- ing, the native chunks are encoded by linear combina- tions to form code chunks. The native and code chunks are distributed over n > k clouds. When an MDS code is used, the original file object may be reconstructed from the chunks contained in any k of the n clouds. Thus, it tolerates the failure of any n − k clouds. We call this feature the MDS property. The extra feature of F-MSR is that reconstructing a single native or code chunk may be achieved by reading up to 50% less data from the sur- viving clouds than reconstructing the whole file. This paper considers a multiple-cloud setting that is double-fault tolerant (e.g., RAID-6) and provides data availability toward at most two cloud failures (e.g., a few days of outages [7]). That is, we set k = n − 2. We expect that such a fault tolerance level suffices in prac- tice. Given that a permanent failure is less frequent but possible, our primary objective is to minimize the cost of storage repair for a permanent single-cloud failure, due to the migration of data over the clouds. We define the repair traffic as the amount of outbound data being read from other surviving clouds during the single-cloud failure recovery. Our goal is to minimize the repair traffic for cost-effective repair. Here, we do not consider the inbound traffic (i.e., the data being written to a cloud), as it is free of charge in many cloud vendors (see Table 1 in Section 5). We now show how F-MSR saves the repair traffic via an example. Suppose that we store a file of size M on four clouds, each viewed as a logical storage node. Let us first consider RAID-6, which is double-fault tolerant. Here, we consider the RAID-6 implementation based on Reed-Solomon codes [26], as shown in Figure 2(a). We divide the file into two native chunks (i.e., A and B) of size M/2 each. We add two code chunks formed by the linear combinations of the native chunks. Suppose now that Node 1 is down. Then the proxy must download the same number of chunks as the original file from two other nodes (e.g., B and A + B from Nodes 2 and 3, re- spectively). It then reconstructs and stores the lost chunk A on the new node. The total storage size is 2M, while the repair traffic is M. We now consider the double-fault tolerant implemen- tation of F-MSR in a proxy-based setting, as shown in Figure 2(b). F-MSR divides the file into four na- tive chunks, and constructs eight distinct code chunks P1, · · · , P8 formed by different linear combinations of the native chunks. Each code chunk has the same size M/4 as a native chunk. Any two nodes can be used to re- cover the original four native chunks. Suppose Node 1 is down. The proxy collects one code chunk from each sur- viving node, so it downloads three code chunks of size M/4 each. Then the proxy regenerates two code chunks P1 and P2 formed by different linear combinations of the three code chunks. Note that P1 and P2 are still lin- ear combinations of the native chunks. The proxy then writes P1 and P2 to the new node. In F-MSR, the stor- age size is 2M (as in RAID-6), but the repair traffic is 0.75M, which is 25% of saving. To generalize F-MSR for n storage nodes, we divide a file of size M into 2(n − 2) native chunks, and generate 4(n − 2) code chunks. Then each node will store two code chunks of size M 2(n−2) each. Thus, the total storage size is Mn n−2 . To repair a failed node, we download one chunk from each of n − 1 nodes, so the repair traffic is M(n−1) 2(n−2) . In contrast, for RAID-6, the total storage size is also Mn n−2 , while the repair traffic is M. When n is large, F-MSR can save the repair traffic by close to 50%. Note that F-MSR keeps only code chunks rather than native chunks. To access a single chunk of a file, we need to download and decode the entire file for the particular chunk. Nevertheless, F-MSR is acceptable for long-term archival applications, whose read frequency is typically low [6]. Also, to restore backups, it is natural to retrieve the entire file rather than a particular chunk. This paper considers the baseline RAID-6 implemen- tation using Reed-Solomon codes. Its repair method is to reconstruct the whole file, and is applicable for all era- sure codes in general. Recent studies [18, 28, 29] show that data reads can be minimized specifically for XOR- based erasure codes. For example, in RAID-6, data reads can be reduced by 25% compared to reconstructing the whole file [28, 29]. Although such approaches are sub- optimal (recall that F-MSR can save up to 50% of repair traffic in RAID-6), their use of efficient XOR operations can be of practical interest.
- 3. ANode 1 Proxy BNode 2 A+BNode 3 A+2BNode 4 A B Object of size M B A+B A New node A Node 1 Proxy Node 2 Node 3 Node 4 Object of size M New node A B C D P1 P2 P3 P4 P7 P8 P5 P6 P3 P5 P7 P1’ P2’ P1’ P2’ (a) RAID-6 (b) F-MSR Figure 2: Examples of repair operations in RAID-6 and F-MSR with n = 4 and k = 2. 3 F-MSR Implementation In this section, we present a systematic approach for im- plementing F-MSR. We specify three operations for F- MSR on a particular file object: (1) file upload; (2) file download; (3) repair. A key difference of our imple- mentation from prior theoretical studies is that we do not require storage nodes to have encoding capabilities, so our implementation can be compatible with today’s cloud storage. Another key design issue is that instead of simply generating random linear combinations for code chunks (as assumed in [8]), we also guarantee that the generated linear combinations always satisfy the MDS property to ensure data availability, even after iterative repairs. Here, we implement F-MSR as an MDS code for general (n,k). We assume that each cloud repository corresponds to a logical storage node. 3.1 File Upload To upload a file F, we first divide it into k(n − k) equal- size native chunks, denoted by (Fi)i=1,2,···,k(n−k). We then encode these k(n − k) native chunks into n(n − k) code chunks, denoted by (Pi)i=1,2,···,n(n−k). Each Pi is formed by a linear combination of the k(n − k) na- tive chunks. Specifically, we let EM = [αi,j] be an n(n−k)×k(n−k) encoding matrix for some coefficients αij (where i = 1, . . . , n(n−k) and j = 1, . . . , k(n−k)) in the Galois field GF(28 ). We call a row vector of EM an encoding coefficient vector (ECV), which con- tains k(n − k) elements. We let ECVi denote the ith row vector of EM. We compute each Pi by the scalar product of ECVi and the native chunk vector (Fi), i.e., Pi = k(n−k) j=1 αijFj for i = 1, 2, · · · , n(n − k), where all arithmetic operations are performed over GF(28 ). The code chunks are then evenly stored in the n storage nodes, each having (n − k) chunks. Also, we store the whole EM in a metadata object that is then replicated to all storage nodes (see Section 4). There are many ways of constructing EM, as long as it satisfies the MDS prop- erty and the repair MDS property (see Section 3.3). Note that the implementation details of the arithmetic opera- tions in Galois Fields are extensively discussed in [15]. 3.2 File Download To download a file, we first download the correspond- ing metadata object that contains the ECVs. Then we select any k of the n storage nodes, and download the k(n − k) code chunks from the k nodes. The ECVs of the k(n−k) code chunks can form a k(n−k)×k(n−k) square matrix. If the MDS property is maintained, then by definition, the inverse of the square matrix must ex- ist. Thus, we multiply the inverse of the square matrix with the code chunks and obtain the original k(n − k) native chunks. The idea is that we treat F-MSR as a stan- dard Reed-Solomon code, and our technique of creating an inverse matrix to decode the original data has been described in the tutorial [22]. 3.3 Iterative Repairs We now consider the repair of F-MSR for a file F for a permanent single-node failure. Given that F-MSR regen- erates different chunks in each repair, one challenge is to ensure that the MDS property still holds even after itera- tive repairs. This is in contrast to regenerating the exact lost chunks as in RAID-6, which guarantees the invari- ance of the stored chunks. Here, we propose a two-phase checking heuristic as follows. Suppose that the (r − 1)th repair is successful, and we now consider how to operate the rth repair for a single permanent node failure (where r ≥ 1). We first check if the new set of chunks in all stor- age nodes satisfies the MDS property after the rth repair. In addition, we also check if another new set of chunks in all storage nodes still satisfies the MDS property after the (r + 1)th repair, should another single permanent node failure occur (we call this the repair MDS property). We now describe the rth repair as follows. Step 1: Download the encoding matrix from a surviving node. Recall that the encoding matrix EM specifies the ECVs for constructing all code chunks via linear combi- nations of native chunks. We use these ECVs for our later two-phase checking heuristic. Since we embed EM in a metadata object that is replicated, we can simply down- load the metadata object from one of the surviving nodes. Step 2: Select one random ECV from each of the n − 1 surviving nodes. Note that each ECV in EM corre-
- 4. sponds to a code chunk. We randomly pick one ECV from each of the n − 1 surviving nodes. We call these ECVs to be ECVi1 , ECVi2 , · · ·, ECVin−1 . Step 3: Generate a repair matrix. We construct a (n − k)×(n−1) repair matrix RM = [γi,j], where each ele- ment γi,j (where i = 1, . . . , n−k and j = 1, . . . , n−1) is randomly selected in GF(28 ). Note that the idea of gen- erating a random matrix for reliable storage is consistent with that in [24]. Step 4: Compute the ECVs for the new code chunks and reproduce a new encoding matrix. We multiply RM with the ECVs selected in Step 2 to construct n − k new ECVs, denoted by ECVi = n−1 j=1 γi,jECVij for i = 1, 2, · · · , n − k. Then we reproduce a new encoding matrix, denoted by EM , which is given by: ith row vector of EM = ECVi, i is a surviving node, ECVi, i is a new node. Step 5: Given EM , check if both the MDS and repair MDS properties are satisfied. Intuitively, we verify the MDS property by enumerating all n k subsets of k nodes to see if each of their corresponding encoding matrices forms a full rank. For the repair MDS property, we check that for any failed node (out of n nodes), we can collect any one out of n−k chunks from the other n−1 surviving nodes and reconstruct the chunks in the new node, such that the MDS property is maintained. The number of checks performed for the repair MDS property is at most n(n − k)n−1 n k . If n is small, then the enumeration complexities for both MDS and repair MDS properties are manageable. If either one phase fails, then we return to Step 2 and repeat. We emphasize that Steps 1 to 5 only deal with the ECVs, so their overhead does not depend on the chunk size. Step 6: Download the actual chunk data and regenerate new chunk data. If the two-phase checking in Step 5 succeeds, then we proceed to download the n−1 chunks that correspond to the selected ECVs in Step 2 from the n − 1 surviving storage nodes to NCCloud. Also, using the new ECVs computed in Step 4, we regenerate new chunks and upload them from NCCloud to a new node. Remark: We claim that in addition to checking the MDS property, checking the repair MDS property is essential for iterative repairs. We conduct simulations to justify that checking the repair MDS property can make itera- tive repairs sustainable. In our simulations, we consider multiple rounds of permanent node failures for different values of n. Specifically, in each round, we randomly pick a node to permanently fail and trigger a repair. We say a repair is bad if the loop of Steps 2 to 5 is repeated over 10 times. We observe that without checking the re- pair MDS property, we see a bad repair very quickly, say after no more than 7 and 2 rounds for n = 8 and n = 12, respectively. On the other hand, checking the re- pair MDS property makes iterative repairs sustainable for hundreds of rounds for different values of n, and we do not yet find any bad repair after extensive simulations. 4 NCCloud Design and Implementation We implement NCCloud as a proxy that bridges user applications and multiple clouds. Its design is built on three layers. The file system layer presents NCCloud as a mounted drive, which can thus be easily interfaced with general user applications. The coding layer deals with the encoding and decoding functions. The storage layer deals with read/write requests with different clouds. Each file is associated with a metadata object, which is replicated at each repository. The metadata object holds the file details and the coding information (e.g., encoding coefficients for F-MSR). NCCloud is mainly implemented in Python, while the storage schemes are implemented in C for better effi- ciency. The file system layer is built on FUSE [12]. The coding layer implements both RAID-6 and F-MSR. RAID-6 is built on zfec [30], and our F-MSR implemen- tation mimics the optimizations made in zfec for a fair comparison. Recall that F-MSR generates multiple chunks to be stored on the same repository. To save the request cost overhead (see Table 1), multiple chunks destined for the same repository are aggregated before upload. Thus, F-MSR keeps only one (aggregated) chunk per file ob- ject on each cloud, as in RAID-6. To retrieve a specific chunk, we calculate its offset within the combined chunk and issue a range GET request. 5 Evaluation We now use our NCCloud prototype to evaluate RAID-6 (based on Reed-Solomon codes) and F-MSR in multiple- cloud storage. In particular, we focus on the setting n = 4 and k = 2. We expect that using n = 4 clouds may suffice for practical deployment. Based on this setting, we allow data retrieval with at most two cloud failures. The goal of our experiments is to explore the practi- cality of using F-MSR in multiple-cloud storage. Our evaluation consists of two parts. We first compare the monetary costs of using RAID-6 and F-MSR based on the price plans of today’s cloud vendors. We also em- pirically evaluate the response time performance of our NCCloud prototype atop a local cloud and also a com- mercial cloud vendor. 5.1 Cost Analysis Table 1 shows the monthly price plans for three major vendors as of September 2011. For Amazon S3, we take the cost from the first chargeable usage tier (i.e., storage
- 5. S3 RS Azure Storage (per GB) $0.14 $0.15 $0.15 Data transfer in (per GB) free free free Data transfer out (per GB) $0.12 $0.18 $0.15 PUT,POST (per 10K requests) $0.10 free $0.01 GET (per 10K requests) $0.01 free $0.01 Table 1: Monthly price plans (in US dollars) for Amazon S3 (US Standard), Rackspace Cloud Files and Windows Azure Storage, as of September, 2011. usage within 1TB/month; data transferred out more than 1GB/month but less than 10TB/month). From the analysis in Section 2, we can save 25% of the download traffic during storage repair when n = 4. The storage size and the number of chunks being gen- erated per file object are the same in both RAID-6 and F-MSR (assuming that we aggregate chunks in F-MSR as described in Section 4). However, in the analysis, we have ignored two practical considerations: the size of metadata (Section 4) and the number of requests is- sued during repair. We now argue that they are negligi- ble and that the simplified calculations based only on file size suffice for real-life applications. Metadata size: Our implementation currently keeps the F-MSR metadata size within 160B, regardless of the file size. NCCloud aims at long-term backups (see Sec- tion 2), and can be integrated with other backup applica- tions. Existing backup applications (e.g., [27, 11]) typi- cally aggregate small files into a larger data chunk in or- der to save the processing overhead. For example, the de- fault setting for Cumulus [27] creates chunks of around 4MB each. The metadata size is thus usually negligible. Number of requests: From Table 1, we see that some cloud vendors nowadays charge for requests. RAID-6 and F-MSR differ in the number of requests when re- trieving data during repair. Suppose that we store a file of size 4MB with n = 4 and k = 2. During repair, RAID-6 and F-MSR retrieve two and three chunks, re- spectively (see Figure 2). The cost overhead due to the GET request for RAID-6 is at most 0.427%, and that for F-MSR is at most 0.854%, a mere 0.427% increase. 5.2 Response Time Analysis We deploy our NCCloud prototype in real environments. We then evaluate the response time performance of dif- ferent operations in two scenarios. The first part ana- lyzes in detail the time taken by different NCCloud op- erations, and is done on a local cloud storage in order to lessen the effects of network fluctuations. The second part evaluates how NCCloud actually performs in com- mercial clouds. All results are averaged over 40 runs. 0 10 20 30 40 50 500 400 300 200 100 50 10 1 Responsetime (seconds) File size (MB) RAID-6 F-MSR (a) File upload 0 2 4 6 8 10 12 500 400 300 200 100 50 10 1 Responsetime (seconds) File size (MB) RAID-6 F-MSR (b) File download 0 5 10 15 20 25 30 35 500 400 300 200 100 50 10 1 Responsetime (seconds) File size (MB) RAID-6 (code chunk repair) RAID-6 (native chunk repair) F-MSR (c) Repair Figure 3: Response times of main NCCloud operations. 0 10 20 30 40 50 60 RAID-6 F-MSR RAID-6 F-MSR R6(code) R6(native) F-MSR Timetaken(seconds) (a) Upload (b) Download (c) Repair File I/O Upload Download Encoding Decoding Figure 4: Breakdown of response time when dealing with 500MB file. 5.2.1 On a Local Cloud The experiment on a local cloud is carried out on an object-based storage platform based on OpenStack Swift 1.4.2 [21]. NCCloud is mounted on a machine with Intel Xeon E5620 and 16GB RAM. This machine is connected to an OpenStack Swift platform attached with a number of storage servers, each with Intel Core i5-2400 and 8GB RAM. We create (n+1) = 5 containers on Swift, so each container resembles a cloud repository (one of them is a spare node used in repair). In this experiment, we test the response time of three basic operations of NCCloud: (a) file upload; (b) file download; (c) repair. We use eight randomly generated files from 1MB to 500MB as the data set. We set the path of a chosen repository to a non-existing location to simu- late a node failure in repair. Note that there are two types of repair for RAID-6, depending on whether the failed node contains a native chunk or a code chunk.
- 6. 0 1 2 3 4 5 6 10 5 2 1 Responsetime (seconds) File size (MB) RAID-6 F-MSR (a) File upload 0 0.5 1 1.5 2 2.5 3 10 5 2 1 Responsetime (seconds) File size (MB) RAID-6 F-MSR (b) File download 0 1 2 3 4 5 6 7 10 5 2 1 Responsetime (seconds) File size (MB) RAID-6 (code chunk repair) RAID-6 (native chunk repair) F-MSR (c) Repair Figure 5: Response times of NCCloud on Azure. Figure 3 shows the response times of all three op- erations (with 95% confidence intervals plotted), and Figure 4 shows five key constituents of the response time when dealing with a 500MB file. Figure 3 shows that RAID-6 has less response time in file upload and download. With the help of Figure 4, we pinpoint the overhead of F-MSR over RAID-6. Due to having the same MDS property, RAID-6 and F-MSR exhibit similar data transfer time during upload/download. However, F- MSR displays a noticeable encoding/decoding overhead over RAID-6. When uploading a 500MB file, RAID-6 takes 1.490s to encode while F-MSR takes 5.365s; when downloading a 500MB file, no decoding is needed in the case of RAID-6 as the native chunks are available, but F-MSR takes 2.594s to decode. On the other hand, F-MSR has slightly less response time in repair. The main advantage of F-MSR is that it needs to download less data during repair. In repairing a 500MB file, F-MSR spends 3.887s in download, while the native-chunk repair of RAID-6 spends 4.832s. Although RAID-6 generally has less response time than F-MSR in a local cloud environment, we expect that the encoding/decoding overhead of F-MSR can be easily masked by network fluctuations over the Internet, as will be shown next. 5.2.2 On a Commercial Cloud The following experiment is carried out on a machine with Intel Xeon E5530 and 16GB RAM running 64-bit Ubuntu 9.10. We repeat the three operations in Sec- tion 5.2.1 on four randomly generated files from 1MB to 10MB atop Windows Azure Storage. On Azure, we cre- ate (n+1) = 5 containers to mimic different cloud repos- itories. The same operation for both RAID-6 and F-MSR are run interleaved to lessen the effect of network fluctu- ation on the comparison due to different times of the day. Figure 5 shows the results for different file sizes with 95% confidence intervals plotted. Note that although we have used only Azure in this experiment, actual usage of NCCloud should stripe data over different vendors and locations for better availability guarantees. From Figure 5, we do not see distinct response time differences between RAID-6 and F-MSR in all opera- tions. Furthermore, on the same machine, F-MSR takes around 0.150s to encode and 0.064s to decode a 10MB file (not shown in the figures). These constitute roughly 3% of the total upload and download times (4.962s and 2.240s respectively). Given that the 95% confidence in- tervals for the upload and download times are 0.550s and 0.438s respectively, network fluctuation plays a big- ger role in determining the response time. Overall, we demonstrate that F-MSR does not have significant per- formance overhead over our baseline RAID-6 implemen- tation. 6 Related Work There are several systems proposed for multiple-cloud storage. HAIL [5] provides integrity and availabil- ity guarantees for stored data. DEPSKY [4] addresses Byzantine Fault Tolerance by combining encryption and erasure coding for stored data. RACS [1] uses erasure coding to mitigate vendor lock-ins when switching cloud vendors. It retrieves data from the cloud that is about to be failed and moves the data to the new cloud. Un- like RACS, NCCloud excludes the failed cloud in repair. All the above systems are based on erasure codes, while NCCloud considers regenerating codes with an emphasis on storage repair. Regenerating codes (see survey [9]) exploit the opti- mal trade-off between storage cost and repair traffic. Ex- isting studies mainly focus on theoretical analysis. Sev- eral studies (e.g., [10, 13, 19]) empirically evaluate ran- dom linear codes for peer-to-peer storage. However, their evaluations are mainly based on simulations. NCFS [17] implements regenerating codes, but does not con- sider MSR codes that are based on linear combinations. Here, we consider the F-MSR implementation, and per- form empirical experiments in multiple-cloud storage. 7 Conclusions We present NCCloud, a multiple-cloud storage file sys- tem that practically addresses the reliability of today’s cloud storage. NCCloud not only achieves fault tolerance of storage, but also allows cost-effective repair when a cloud permanently fails. NCCloud implements a practi- cal version of the functional minimum storage regenerat- ing code (F-MSR), which regenerates new chunks during
- 7. repair subject to the required degree of data redundancy. Our NCCloud prototype shows the effectiveness of F- MSR in accessing data, in terms of monetary costs and response times. The source code of NCCloud is available at http://ansrlab.cse.cuhk.edu.hk/software/nccloud. 8 Acknowledgments We would like to thank our shepherd, James Plank, and the anonymous reviewers for their valuable comments. This work was supported by grant AoE/E-02/08 from the University Grants Committee of Hong Kong. References [1] H. Abu-Libdeh, L. Princehouse, and H. Weather- spoon. RACS: A Case for Cloud Storage Diversity. In Proc. of ACM SoCC, 2010. [2] R. Ahlswede, N. Cai, S.-Y. R. Li, and R. W. Ye- ung. Network Information Flow. IEEE Trans. on Information Theory, 46(4):1204–1216, Jul 2000. [3] M. Armbrust, A. Fox, R. Griffith, A. D. Joseph, R. Katz, A. Konwinski, G. Lee, D. Patterson, A. Rabkin, I. Stoica, and M. Zaharia. A View of Cloud Computing. Communications of the ACM, 53(4):50–58, 2010. [4] A. Bessani, M. Correia, B. Quaresma, F. Andr´e, and P. Sousa. DEPSKY: Dependable and Secure Storage in a Cloud-of-Clouds. In Proc. of ACM Eu- roSys, 2011. [5] K. D. Bowers, A. Juels, and A. Oprea. HAIL: A High-Availability and Integrity Layer for Cloud Storage. In Proc. of ACM CCS, 2009. [6] B. Chen, R. Curtmola, G. Ateniese, and R. Burns. Remote Data Checking for Network Coding-Based Distributed Storage Systems. In Proc. of ACM CCSW, 2010. [7] CNNMoney. Amazon’s cloud is back, but still hazy. http://money.cnn.com/2011/04/25/ technology/amazon_cloud/index.htm. [8] A. G. Dimakis, P. B. Godfrey, Y. Wu, M. Wain- wright, and K. Ramchandran. Network Coding for Distributed Storage Systems. IEEE Trans. on In- formation Theory, 56(9):4539–4551, Sep 2010. [9] A. G. Dimakis, K. Ramchandran, Y. Wu, and C. Suh. A Survey on Network Codes for Dis- tributed Storage. Proc. of the IEEE, 99(3):476–489, Mar 2011. [10] A. Duminuco and E. Biersack. A Practical Study of Regenerating Codes for Peer-to-Peer Backup Sys- tems. In Proc. of IEEE ICDCS, 2009. [11] B. Escoto and K. Loafman. Duplicity. http:// duplicity.nongnu.org/. [12] FUSE. http://fuse.sourceforge.net/. [13] C. Gkantsidis and P. Rodriguez. Network coding for large scale content distribution. In Proc. of IN- FOCOM, 2005. [14] GmailBlog. Gmail back soon for everyone. http: //gmailblog.blogspot.com/2011/02/ gmail-back-soon-for-everyone.html. [15] K. M. Greenan, E. L. Miller, and T. J. E. Schwarz. Optimizing Galois Field Arithmetic for Diverse Processor Architectures and Applications. In Proc. of IEEE MASCOTS, 2008. [16] Y. Hu, Y. Xu, X. Wang, C. Zhan, and P. Li. Coop- erative recovery of distributed storage systems from multiple losses with network coding. IEEE JSAC, 28(2):268–276, Feb 2010. [17] Y. Hu, C.-M. Yu, Y.-K. Li, P. P. C. Lee, and J. C. S. Lui. NCFS: On the Practicality and Extensibility of a Network-Coding-Based Distributed File System. In Proc. of NetCod, 2011. [18] O. Khan, R. Burns, J. Plank, and C. Huang. In Search of I/O-Optimal Recovery from Disk Fail- ures. In USENIX HotStorage, 2011. [19] M. Martal´o, M. Picone, M. Amoretti, G. Ferrari, and R. Raheli. Randomized Network Coding in Distributed Storage Systems with Layered Overlay. In Information Theory and Application Workshop, 2011. [20] E. Naone. Are We Safeguarding Social Data? http://www.technologyreview.com/ blog/editors/22924/, Feb 2009. [21] OpenStack Object Storage. http://www. openstack.org/projects/storage/. [22] J. S. Plank. A Tutorial on Reed-Solomon Coding for Fault-Tolerance in RAID-like Systems. Soft- ware - Practice & Experience, 27(9):995–1012, Sep 1997. [23] C. Preimesberger. Many data centers unpre- pared for disasters: Industry group, Mar 2011. http://www.eweek.com/c/a/IT- Management/Many-Data-Centers- Unprepared-for-Disasters-Industry- Group-772367/. [24] M. O. Rabin. Efficient Dispersal of Information for Security, Load Balancing, and Fault Tolerance. Journal of the ACM, 36(2):335–348, Apr 1989. [25] K. V. Rashmi, N. B. Shah, P. V. Kumar, and K. Ramchandran. Explicit Construction of Optimal Exact Regenerating Codes for Distributed Storage. In Proc. of Allerton Conference, 2009.
- 8. [26] I. Reed and G. Solomon. Polynomial Codes over Certain Finite Fields. Journal of the Society for In- dustrial and Applied Mathematics, 8(2):300–304, 1960. [27] M. Vrable, S. Savage, and G. Voelker. Cumu- lus: Filesystem backup to the cloud. In Proc. of USENIX FAST, 2009. [28] Z. Wang, A. Dimakis, and J. Bruck. Rebuilding for Array Codes in Distributed Storage Systems. In IEEE GLOBECOM Workshops, 2010. [29] L. Xiang, Y. Xu, J. Lui, Q. Chang, Y. Pan, and R. Li. A Hybrid Approach to Failed Disk Recovery Using RAID-6 Codes: Algorithms and Performance Eval- uation. ACM Trans. on Storage, 7(3):11, 2011. [30] zfec. http://pypi.python.org/pypi/ zfec.