Performance analysis with_ceph

Performance Analysis with ceph
雲儲存性能分析
Alex Lau 劉俊賢
Software Consultant 研發工程師/顧問
(AvengerMoJo) alau@suse.com

Agenda 議程
SES5 is base on Luminous – The Why? 為何分析性能?
Ceph performance – The How? 如何分析性能?
Ceph analysis – The What? 分析結果是怎樣?
The Future? 大家有什麼展望?

SES5 is base on Luminous – The Why?
為何分析性能?

SUSE Enterprise Storage 5 base on Ceph Luminous
4
• Support x86-64 and AArch64
• Easy to use WebUI openATTIC 3.x
• Simple DeepSea Cluster orchestration
• BlueStore ready with data compression
• Cephfs and NFS-Ganesha ready

Ceph Info 基本資料
Code Developers
782
Core Regular Casual
22 53 705
Total downloads
160,015,454
Unique downloads
21,264,047
Client Servers
(Windows, Linux, Unix)
RADOS (Common Object Store)
Block Devices
Server
Object Storage File Interface
Storage
Server
Storage
Server
Storage
Server
Storage
Server
Server
Server
Storage
Server
Storage
Server
Applications File Share
OSD OSD OSD OSD OSD OSD
NetworkCluster
MON
RBD
iSCSI
S3
SWIFT
CephFS*
Monitors
MONMON
5

SUSE Enterprise Storage其實做了什麼?
基於Ceph
2013: SUSE 加入 Ceph 社區
2015.01 : SUSE Enterprise Storage 1.0
2015.11: SUSE Enterprise Storage 2.0
2016: SUSE 收购 IT-Novum（主要開發存儲管理工具 openATTIC）
2016.09: SUSE 合併 HPE 軟體
2017.09: SUSE Enterprise Storage 5.0 (Beta Now)
6

7
SUSE Enterprise Storage
Roadmap
2016 2017 2018
V3
V4
V5
SUSE Enterprise Storage 4 SUSE Enterprise Storage 5*
Built On
• Ceph Jewel release
• SLES 12 SP 2 (Server)
Manageability
• SES openATTIC management
• Initial Salt integration DeepSea
Interoperability
• AArch64
• CephFS (production use cases)
• NFS Ganesha (Tech Preview)
• NFS access to S3 buckets (Tech Preview)
• CIFS Samba (Tech Preview)
• RDMA/Infiniband (Tech Preview)
Availability
• Multisite object replication
• Asynchronous block mirroring
Built On
• Ceph Luminous release
• SLES 12 SP 3 (Server)
Manageability
• SES openATTIC management phase 2
• SUSE Manager integration
Interoperability
• NFS Ganesha
• NFS access to S3 buckets
Availability
• Asynchronous block mirroring
• Erasure coded block pool
Efficiency
• BlueStore back-end
• Data compression

SUSE 和 Ceph社區的關系
2013年建立Ceph開發團隊,正式加入Ceph的代碼支持
SUSE 是 Ceph理事會8大理事會員之一
一直以來代碼貢獻頭3名, 上次v12 Luminous發報第2多, 單個人第1 Ricardo
1. 10648 Ricardo Dias rdias@suse.com
2. 6422 Sage Weil sweil@redhat.com
有收到 Inktank 和 Mirantis 開發人員
加大研發投入: 2016年中收購存儲管理工具: it-novum (openATTIC)
8

Customer needs! 客戶的需要!
• How high do we need?
• How much do we want to pay?
• Are we counting 4K Random Write?
• Do we need Erasure Coding?
• Can we have the number on BlueStore
vs FileStore?
• What about cephfs over nfs?
• ……
What’s our cluster IOPS?
• 我們需要多高的IOPS?
• 我們想花多少錢?
• 我們只需要看4K隨機寫?
• 我們有糾刪碼的需要嗎?
• 我們能看看BlueStore對比FileStore嗎?
• 我們能看看cephfs上的NFS能怎樣?
• …….
我們的存儲叢集能跑多少IOPS?
9

System Engineer needs! 售前售後工程師的需要!
Putting them here is already show how much we care!
10

The DevOps Needs!開發運維的需要!
Can we have some base line number?
Can we have tools to re-do performance analysis?
BTW we need all that to our live cluster!
能否告訴正常時叢集該跑多快?
能否每天再測試報告叢集現在的性能?
不是分開測而是實際運作中的叢集性能哦!
12

The Developers Needs! 開發者的需要!
How much is BlueStore vs Filestore 4K, 16K, 64K, 1M, 4M write seq, random, read
write performance in hdd vs ssd vs nvme pure and mixed pool, when an OSD is
down, and the recovery process is started and meanwhile tracking how much
memory needed when the capacity are 50% full without stopping the client keep
writing to the cluster….
In short … very complicated analysis
…. 不寫了.. 很多很多要求就是.
p.s. 大家有硬體提供我們測試請找我, 會後我請你… 
13

The Community Project Need! 社區的性能有關項目!
Ceph Brag
• It require to use CBT
- https://github.com/ceph/cbt
• Setting up the CBT environment could
be complicated
• It require another set of multiple nodes
setup
• It is intrusive test that can’t be operate in
a live cluster
• http://ceph.com/performance/
Ceph Brag
• CBT是現在測試的標準
-https://github.com/ceph/cbt
• 設定CBT環境並不簡單
• 需要多台機器誰行測試
• 影響到內容不能向運作中的系統做測試
• 每週的性能有關會
-http://ceph.com/performance/
14

The Goal
• Discover hardware, driver issue from
time to time to check for degrade.
• Intrusive / Non-intrusive performance
testing to do before and after the cluster
being setup
• Able to perform to live and production
cluster without stopping operation
• Low dependence, flexible, open source,
and able to share eventually and align
wth Ceph Brag
• Dynamic using different tools to do
monitoring and tracing but able to
orchestrated by a central admin that fit
the container and cloud story
• 可以不定時發現硬體或驅動出現問題
• 在叢集設定前可以具破壞性的測試但在
運作中的叢集可以用一樣的工具去做同
類但不具破壞性的做法
• 測試可以在正在運作中的系統上進行而
不影響一般運作
• 不用依賴太多三方面的需要, 可以靈活使
用, 開源, 最終可以和Ceph Brag共享一樣
的發報方法
• 可以使用不同的管理和跟蹤工具, 同時可
以利用中央管理工具預備可以和容器和
雲管理整合
15

Ceph Performance – The How?
如何分析性能?
16

Hardware Testing
Network:
ping –c1 –q –W1
Specific for Jumbo Frame
ping –Mdo –s8972 –c1 –q –W1
Bandwidth:
iperf3 –fm –A0 –t10 –c<server ip> -p<port>
Harddisk:
Read:
hdparm –t <osd_partition>
Write:
dd if=/dev/zero of=`mount <osd_data_partition>/test/
conv=fdatasync bs=4K count=10000
17

Cluster Testing
OSD Bench (1G write with 4M block default)
ceph tell osd.0 bench
RADOS Bench
rados –p <pool> <time> <write,seq,rand> -t <thread> --no-cleanup
RBD bench
rbd –p <pool> bench-write <image> --io-size <e.g 4096>
--io-threads <1,4,16 etc> --io-total<total size
e.g.209715200>
--io-pattern <rand | seq>
New version has io-type
rbd -p <pool> bench --io-type read rbd_test --io-size 4096
--io-threads 1 --io-total 209715200 --io-pattern rand
18

Client Testing
FIO can be use against any of the following:
fio --ioengine=libaio --iodepth=32 --direct=1 --rw=write
--bs=4K --size=200MB --filename=/mnt/200M_data --numjobs=1
--name=cephfs_seqwrite_4K --output=cephfs_write_seq_4K.out
RBD Block
--ioengine can be either librbd or mount using libaio as default
ISCSI export
we can use it in windows or other OS that support iscsi
CephFS mount
mount.cephfs mon1,mon2,mon3:/ /mnt/
NFS/CIFS mount
mount.nfs nfs-server-ip:/ /mnt/
19

Lttng Testing
Kernel Tracing make easy!
lttng create –o .
lttng enable-channel --num-subbuf 16 --subbuf-size 8M -k c0
lttng enable-event --kernel --all
lttng enable-event --syscall -a -k -c c0
lttng start
<do something>
lttng stop
lttng destroy
20

Mixing with Salt
It is already in next version of DeepSea
Network:
salt-run net.ping exclude=<non-cluster-ip>
salt-run net.jumbo_ping exclude=<non-cluster-ip>
Bandwidth:
salt-run net.iperf cluster=ceph output=full
salt-run net.iperf exclude=<non-cluster-ip>
21

Experiment
https://github.com/AvengerMoJo/Ceph-Saltstack
Since some of those test could be intrusive to OSD it is not in DeepSea yet.
salt “*” ceph_sles.disk_info
salt “*” ceph_slse.bench_disk /dev/sdb /dev/sdc /dev/sde
( dd write to a mount point a partition can mount to )
22

Lttng ust user space tracing
Client-side tracing only with librbd
compile ceph librbd with using -finstrument-functions
compile ceph with -DWITH_LTTNG=ON
export LD_PRELOAD=/usr/lib64/liblttng-ust-cyg-profile.so
Calling rbd to do something < write >
Stop tracing and collect the result for reporting
Can use Salt to do a multiple nodes tracing and collect all the report for you
23

Collecting Reporting with Salt
Calling salt to start lttng in every nodes of the cluster
salt-run lttng.run
cmd="[fio --ioengine=libaio --iodepth=256 --direct=1
--rw=write --bs=4K --size=20MB
--filename=/mnt/cephfs/20M_data --numjobs=1
–name=write_4K --output=/tmp/cephfs/write_seq_4K.out]"
cmd_server=salt-master
Then collecting all the result back to the master.
24

Ceph Analysis – The What?
分析結是怎樣?
25

Example we use in the following hardware from SUSE
Enterprise Storage Partners
HPE - OEM
- Apollo – Apollo 4200 * 3
• osd with 18 hdd 2 ssd
- Proliant – DL380/DL360 * 3
• mon

Network Ping
Regular Ping:
ping –c1 –q –W1 (rtt higher then 1ms maybe something wrong)
--- 192.168.128.1 ping statistics ---
1 packets transmitted, 1 received, 0% packet loss, time 0ms
rtt min/avg/max/mdev = 0.298/0.298/0.298/0.000 ms
Jumbo Frame:
ping –Mdo –s8972 –c1 –q –W1 ( double normal ping)
--- 192.168.128.5 ping statistics ---
1 packets transmitted, 1 received, 0% packet loss, time 0ms
rtt min/avg/max/mdev = 0.649/0.649/0.649/0.000 ms
27

Network Bandwidth
Bandwidth: iperf3 –fm –A0 –t10 –c<server ip> -p<port>
Connecting to host 192.168.128.1, port 5201
[ 4] local 192.168.128.77 port 41008 connected to 192.168.128.1 port 5201
[ ID] Interval Transfer Bandwidth Retr Cwnd
[ 4] 0.00-1.00 sec 119 MBytes 1002 Mbits/sec 0 368 KBytes
- - - - - - - - - - - - - - - - - - - - - - - - -
[ ID] Interval Transfer Bandwidth Retr
[ 4] 0.00-10.00 sec 1.16 GBytes 992 Mbits/sec 0 sender
[ 4] 0.00-10.00 sec 1.15 GBytes 991 Mbits/sec receiver
28

DeepSea Network Test
salt-run net.ping exclude=<non-cluster-ip>
Succeeded: 8 addresses from 7 minions average rtt 0.15 ms
salt-run net.jumbo_ping exclude=<non-cluster-ip>
Succeeded: 8 addresses from 7 minions average rtt 0.26 ms
salt-run net.iperf exclude=192.168.128.9 cluster=ceph output=full
29
192.168.128.1:
8644.0 Mbits/sec
192.168.128.2:
10360.0 Mbits/sec
192.168.128.3:
9336.0 Mbits/sec
192.168.128.4:
9588.56 Mbits/sec
192.168.128.5:
10187.0 Mbits/sec
192.168.128.6:
10465.0 Mbits/sec

Disk Direct Read Write
Read: hdparm –t
HDD: Timing buffered disk reads: 618 MB in 3.01 seconds = 205.58 MB/sec
SSD: Timing buffered disk reads: 1510 MB in 3.00 seconds = 503.25 MB/sec
Write: Direct dd right zero into disk mount point
/usr/bin/dd if=/dev/zero of=/var/lib/ceph/osd/ceph-0/test conv=fdatasync
bs=4K count=10000
10000+0 records in
10000+0 records out
HDD: 40960000 bytes (41 MB, 39 MiB) copied, 0.257385 s, 159 MB/s
SSD: 40960000 bytes (41 MB, 39 MiB) copied, 0.117944 s, 347 MB/s
With RAID Card Cache Enable
HDD: 40960000 bytes (41 MB, 39 MiB) copied, 0.038911 s, 1.1 GB/s
30

OSD Bench (without external communication)
OSD Bench (1G write with 4M block default)
ceph tell osd.<num> bench
SSD
{
"bytes_written": 1073741824, (1G)
"blocksize": 4194304, (4M)
"bytes_per_sec": 260063627 (248M)
}
RAID 0 HDD with cache
{
"bytes_written": 1073741824,
"blocksize": 4194304,
"bytes_per_sec": 464233957 (442M)
}
31

RADOS Bench: rados –p hdd3 bench 20 write -t 32 –b 4096 --no-cleanup
Maintaining 32 concurrent writes of 4096 bytes to objects of size 4096 for up
to 20 seconds or 0 objects
2017-07-11 13:59:02.450884 min lat: 0.000861743 max lat: 0.193705 avg lat:
0.0033328
sec Cur ops started finished avg MB/s cur MB/s last lat(s) avg lat(s)
20 9 191944 191935 37.4835 36.7539 0.00325612 0.0033328
Total time run: 20.008286
Total writes made: 191944
Write size: 4096
Object size: 4096
Bandwidth (MB/sec): 37.4735
Stddev Bandwidth: 2.88217
Max bandwidth (MB/sec): 44.332
Min bandwidth (MB/sec): 32.9258
Average IOPS: 9593
Stddev IOPS: 737
Max IOPS: 11349
Min IOPS: 8429
Average Latency(s): 0.00333329
Stddev Latency(s): 0.00611436
Max latency(s): 0.193705
Min latency(s): 0.000861743
Byte Size = 4M
Total time run: 20.186976
Total writes made: 4985
Write size: 4194304
Object size: 4194304
Bandwidth (MB/sec): 987.766
Stddev Bandwidth: 178.349
Max bandwidth (MB/sec): 1200
Min bandwidth (MB/sec): 596
Average IOPS: 246
Stddev IOPS: 44
Max IOPS: 300
Min IOPS: 149
Average Latency(s): 0.129235
Stddev Latency(s): 0.228253
Max latency(s): 6.50408
Min latency(s): 0.0230147
32

RBD Bench: rbd write with seq and rand
rbd –p ssd bench-write rbd_test --io-size 4096 --io-threads 16
--io-total 209715200 --io-pattern seq
elapsed: 2 ops: 51200 ops/sec: 17947.54 bytes/sec: 73513129.85
rbd –p hdd bench-write rbd_test --io-size 4096 --io-threads 16
--io-total 209715200 --io-pattern rand
HDD with RAID card cache run faster in Sequence compare to Random
33

RBD Feature enable vs disable
To enable all the features in rbd: striping, object-map, fast-diff, deep-flatten, journaling
HDD with RAID cache run much worst in Random
34

RBD Block XFS mount hdd (cache) 4K IOPS benchmark
fio --ioengine=libaio --iodepth=256 --direct=1 --rw=write --bs=4K --size=200MB
--filename=/mnt/hdd/200M_data --numjobs=1 –name=seq_write_4K
write: io=204800KB, bw=17667KB/s, iops=4416, runt= 11592msec
fio --ioengine=libaio --iodepth=256 --direct=1 --rw=randwrite --bs=4K --size=200MB
--filename=/mnt/hdd/200M_data --numjobs=1 --name=write_4K
fio --ioengine=libaio --iodepth=256 --direct=1 --rw=read --bs=4K --size=200MB
--filename=/mnt/hdd/200M_data --numjobs=1 --name=read_4K
read : io=204800KB, bw=71235KB/s, iops=17808, runt= 2875msec
fio --ioengine=libaio --iodepth=256 --direct=1 --rw=randread --bs=4K --size=200MB
--filename=/mnt/hdd/200M_data --numjobs=1 --name=read_4K
35

RBD Block XFS mount ssd 4K IOPS benchmark
fio --ioengine=libaio --iodepth=256 --direct=1 --rw=write --bs=4K --size=200MB
--filename=/mnt/ssd/200M_data --numjobs=1 –name=seq_write_4K
fio --ioengine=libaio --iodepth=256 --direct=1 --rw=randwrite --bs=4K --
size=200MB
--filename=/mnt/ssd/200M_data --numjobs=1 --name=write_4K
fio --ioengine=libaio --iodepth=256 --direct=1 --rw=read --bs=4K --size=200MB
--filename=/mnt/ssd/200M_data --numjobs=1 --name=read_4K
fio --ioengine=libaio --iodepth=256 --direct=1 --rw=randread --bs=4K --size=200MB
--filename=/mnt/ssd/200M_data --numjobs=1 --name=read_4K
36

RBD Block XFS hdd (cache) 1M ThoughtPut benchmark
fio --ioengine=libaio --iodepth=16 --direct=1 --rw=write –bs=1M --size=4GB
--filename=/mnt/hdd/4G_data --numjobs=1 –name=seq_write_1M
write: io=4096.0MB, bw=1043.6MB/s, iops=1043, runt= 3925msec
fio --ioengine=libaio –iodepth=16 --direct=1 --rw=randwrite --bs=1M --size=4GB
--filename=/mnt/hdd/4G_data --numjobs=1 --name=write_1M
write: io=4096.0MB, bw=1146.6MB/s, iops=1146, runt= 3574msec
fio --ioengine=libaio --iodepth=16 --direct=1 --rw=read --bs=1M –size=4GB
--filename=/mnt/hdd/4G_data --numjobs=1 --name=read_1M
read : io=4096.0MB, bw=316790KB/s, iops=309, runt= 13240msec
fio --ioengine=libaio --iodepth=16 --direct=1 --rw=randread --bs=1M –size=4GB
--filename=/mnt/hdd/4G_data --numjobs=1 --name=read_1M
37

RBD Block XFS ssd 1M ThoughtPut benchmark
fio --ioengine=libaio --iodepth=16 --direct=1 --rw=write --bs=1M --size=4GB
--filename=/mnt/ssd/4G_data --numjobs=1 –name=seq_write_4G
write: io=4096.0MB, bw=540225KB/s, iops=527, runt= 7764msec
fio --ioengine=libaio –iodepth=16 --direct=1 --rw=randwrite --bs=1M --size=4GB
--filename=/mnt/ssd/4G_data --numjobs=1 --name=write_4G
write: io=4096.0MB, bw=554142KB/s, iops=541, runt= 7569msec
fio --ioengine=libaio --iodepth=16 --direct=1 --rw=read --bs=1M --size=4GB
--filename=/mnt/ssd/4G_data --numjobs=1 --name=read_4G
fio --ioengine=libaio --iodepth=16 --direct=1 --rw=randread --bs=1M --size=4GB
--filename=/mnt/ssd/4G_data --numjobs=1 --name=read_4G
38

Lttng tracing regular RBD, Journaling enable RBD
Default:
SEC OPS OPS/SEC BYTES/SEC
Journaling:
bench type write io_size 4 io_threads 1 bytes 4 pattern random
SEC OPS OPS/SEC BYTES/SEC
39

Using Tracecompress
Eclipse base UI, easy to use, cross platform (java)
http://tracecompass.org/
40

Regular Layering RBD “tp_librbd”
41

RBD with remote Journaling “tp_librbd_journ”
42

https://docs.google.com/drawings/d/1s9pRVygdPzBtQyYrvPueF8Cm6Xi1d4oy5gNMOjx2Wzo
43
[…] Waiting on lock()
AioCompletion() Aio_write2()
rbd
Copy bufferlist aio_write()
ImageWriteRequest()
::queue()
tp_librbd
ThreadPoolPointer::void_process()
ImageWriteRequest()
append_journal_event()
Send_object_cache_request ()

https://docs.google.com/drawings/d/1s9pRVygdPzBtQyYrvPueF8Cm6Xi1d4oy5gNMOjx2Wzo
44
Waiting on lock()
tp_librbd
ThreadPoolPointer::void_process()
ImageWriteRequest()
append_journal_event()
Send_object_cache_request ()
_void_process()
FunctionContext::finish()
ObjectWriteOperation::append()
IoCtx::aio_operate()
Waiting on lock()
Context::complete()fn_anonymous
tp_librbd_journ

The Future?
以後的展望?
46

The Future
• Setup partner with production hardware in community that can run the non-
intrusive test case in regular bases to compare with different ceph cluster.
• Include more client side multiple nodes tracing ( container / cloud ready )
• Filestore and Bluestore with detail features on and off
• More in-depth ARM AArch64 ceph performance report
• Upstream is already doing a lot of recovery related to timing issue
• Chinese Translation related to performance analysis tools and method
• 提供服務給社區運作中的叢集非破壞式的測試方法, 給大家參考和對比差異
• 提供更完整的多點跟蹤功能 (容器內/雲內)
• 針對Filestore和Bluestore做更深人的測試,觀察打開不同功能後的性能比較
• 再深人去了解ARM64運作ceph時的性能報告
• 上游正著手觀察Recovery有關的不同選項做分析
• 翻譯有關這麼多不同性能測試工具和方法的文檔
47

Performance analysis with_ceph

More Related Content

What's hot (19)

Similar to Performance analysis with_ceph (20)

More from Alex Lau (8)

Recently uploaded (20)

Performance analysis with_ceph

Editor's Notes