GraphLab

A presentation by Tushar Sudhakar Jee
A Distributed Framework for Machine Learning and Data Mining in the Cloud

Bulk Synchronous Parallel(BSP)
• A bridging model for designing parallel
Algorithms(eg: message relaying).
• Implemented by Google Pregel 2010.
• The model consists of :
1. Concurrent computation:Every
participating processor may
perform local computations.
2. Communication: The processes
exchange data between
themselves to facilitate remote
data storage capabilities.
3. Barrier synchronisation : When a
process reaches this point (the barrier),
it waits until all other processes have
reached the same barrier.

Bulk Synchronous Parallel(BSP)
• Advantages:
1. No worries about Race conditions.
2. Barrier guarantees Data consistency.
3. Simpler to make fault tolerant,save data on barrier.
• Disadvantages:
1. Costly performance penalties since runtime of each phase is decided by
slowest machine.
2. Fail to support the properties of asynchronous,graph-parallel and dynamic
computation,critical to Machine Learning and Data Mining Community.

Asynchronous processing
• Implemented by GraphLab 2010, 2012 .
• Advantages:
1. Directly targets properties of asynchronous,graph-parallel and dynamic
computation,critical to Machine Learning and Data Mining Community.
2. Involves updating parameters using most recent values as input,most closely
related to Sequential execution.
• Disadvantages:
1. Race conditions can happen all the time.

Why GraphLab?
•Implementing Machine Learning and Data
Mining algorithms in parallel on current systems
like Hadoop,MPI and MapReduce is prohibitively
complex and costly.
•It targets asynchronous, dynamic, graph-
parallel computation in the shared-memory
setting as needed by the MLDM community.

MLDM Algorithm Properties
•Graph Structured Computation
•Asynchronous Iterative Computation
•Dynamic Computation
•Serializability

Graph Structured Computation
•Recent advances in MLDM focus on modeling the dependencies between data, as it
allows extracting more signal from noisy data.
For example, modeling the dependencies between similar shoppers allows us to make
better product recommendations than treating them in isolation.
•Consequently, there has been a recent shift toward graph-parallel abstractions like
Pregel and GraphLab that naturally express Computational dependencies.

Asynchronous Iterative Computation
•Synchronous systems update all parameters
simultaneously (in parallel) using parameter
values from the previous time step as input.
•Asynchronous systems update parameters
using the most recent parameter values as input.
Many MLDM algorithms benefit from
asynchronous systems.

Dynamic Computation
•Dynamic computation allows the
algorithm to save time since it only
recomputes vertices with recently
updated neighbors .
•Static computation requires the
algorithm to update all vertices equally
often. This wastes time recomputing
vertices that have already converged.

Serializability
•Serializability ensures that all parallel
executions have an equivalent sequential
execution,thereby eliminating race
conditions.
•MLDM algorithms converge faster if
serializability is ensured.Gibbs sampling,
requires serializability for correctness.

Distributed GraphLab Abstraction
•Data Graph
•Update function
•Sync Operation
•GraphLab Execution Model
•Ensuring Serializability

Data Graph
•The GraphLab abstraction stores the program
state as a directed graph called the data
graph, G = (V, E, D), where D is the user
defined data.
Data here represents model parameters,
algorithm state, and statistical data.
Data graph

Data Graph (PageRank example):

Update function
• An update function is a stateless procedure
that modifies the data within the scope of a
vertex and schedules the future execution of the
update functions on other vertices.
•The function takes a vertex v and its scope Sv
and returns new versions of the data in the
scope as well as a set vertices T:
Update: f(v, Sv ) -> (Sv ,T)

Update function(PageRank example):
•The update function for PageRank
computes a weighted sum of the
current ranks of neighboring vertices
and assigns it as the rank of the
current vertex.
• The Neighbors are scheduled for
update only if the value of the current
vertex crosses the threshold.

Update function(PageRank example):

Sync Operation
•It is an associative commutative sum defined over all scopes in the graph.
•Supports normalization common in MLDM algorithms.
•Runs continuously in the background to maintain updated estimates of the
global value.
•Ensuring serializability of the sync operation is costly and requires
synchronization.

The GraphLab Execution Model
•The model allows the GraphLab
runtime engine to determine best
order in which to run vertices.
•Since many MLDM algorithms
benefit from prioritization,GraphLab
abstraction allows users to assign
priorities to vertices in T.

Ensuring Serializability
•It implies that for every parallel
execution, there exists a sequential
execution of update functions that would
give the same results.
•For Serializability ensure no overlapping
in scopes of concurrently executing scopes
of update functions.
•The greater the consistency, the lower
the parallelism.

Ensuring Serializability(Full Consistency):
• This model ensures that scopes of concurrently executing update functions do not
overlap.
• Update function has complete read/write access to entire scope.
• Concurrently executing update functions must be at least two vertices apart
limiting parallelism.

Ensuring Serializability(Edge Consistency):
• This model ensures each update function has exclusive read/write access to its
vertex and adjacent edges, but read only access to adjacent vertices.
• Increases parallelism by allowing update functions with slightly overlapping
scopes to run in parallel.

Ensuring Serializability(Vertex Consistency):
• This model provides write access to the central vertex data.
• It allows all update functions to be run in parallel, providing maximum
parallelism.
• It is the least consistent.

Using GraphLab(K-means):
• Using GraphLab Create K-means with dataset from the June 2014
Kaggle competition to classify schizophrenic subjects based on MRI scans.
• The original data consists of two sets of features: functional network
connectivity (FNC) features and source-based morphometry (SBM) features,
incorporated into a single SFrame with SFrame.join.
• Data downloaded from public AWS S3 bucket.

Distributed GraphLab Design
•Distributed Data Graph
•Distributed GraphLab Engines
•Fault tolerance
•System design

Distributed Data Graph
•A graph is partitioned into k parts
where k is much greater than the
number of machines.
•Each part, called an atom is stored
as a separate file on a distributed
storage system(Amazon S3).

Distributed GraphLab Engines:
1. Emulates the GraphLab execution model and is responsible for:
•Executing update functions.
•Executing sync operations.
•Maintaining the set of scheduled vertices T.
•Ensuring serializability with respect to the appropriate consistency model
2. Types:
• Chromatic Engine
• Distributed Locking Engine

•It uses vertex coloring to satisfy the edge
consistency model by executing synchronously
all vertices of the same color in the vertex set T
before proceeding to the next color.
•Full consistency model is satisfied by ensuring
that no vertex shares the same color as any of
its distance two neighbors.
•Vertex consistency model is satisfied by
assigning all vertices the same color.
• It executes the set of scheduled vertices T
partially asynchronously.
Chromatic Engine:
Edge Consistency model using
Chromatic Engine.

Distributed Locking Engine
1. Why use it?
•Chromatic engine does not provide
sufficient scheduling flexibility.
• Chromatic engine presupposes availability
of graph coloring which might not always
be readily available.
2.The Distributed Locking Engine uses mutual
exclusion by associating a readers-writers lock with
each vertex.
3.Vertex consistency is achieved by acquiring a write-
lock on the central vertex of each requested scope.

4. Full consistency is achieved by acquiring write-locks on the central vertex and
all adjacent vertices.
5.Edge consistency is achieved by acquiring a write-lock on the central vertex
and read locks on adjacent vertices.

Distributed Locking Engine(Pipelined locking)
•Each machine maintains a pipeline of
vertices for which locks have been
requested, but not yet fulfilled.
•The pipelining system uses callbacks
instead of readers/writer locks since the
latter would halt the pipeline.
•Pipelining reduces latency by
synchronizing locked data immediately
after a machine completes its local lock.

Chandy-Lamport Snapshot Algorithm

Fault Tolerance
•Using a distributed checkpoint mechanism called Snapshot
Update fault tolerance is introduced in GraphLab.
•Snapshot Update can be deployed synchronously or
asynchronously.
•Asynchronous snapshots are more efficient and guarantee
consistent snapshot under the following conditions:
a)Edge consistency is used on all update functions.
b)Schedule completes before the scope is unlocked.
c)Snapshot Update is prioritized over other updates.

Synchronous snapshot have
the “flatline” whereas
asynchronous snapshots
allow computation to
proceed.

System design
● In the Initialization phase the atom file representation of data graph is
constructed.
● In the GraphLab Execution Phase atom files are assigned to individual
execution engines from the DFS.

System design(Locking Engine Design)
• Partition of distributed graph managed within Local Graph storage.
• Cache used to provide access to remote data.
• Scheduler manages vertices in T assigned to the process.
• Each block makes use of block below it.

Applications
● Netflix Movie Recommendation
● Video Co-segmentation(Coseg)
● Named Entity Recognition(NER)

Netflix Movie Recommendation
● It makes use of collaborative filtering to predict
the movie ratings for each user based on the
ratings of similar users.
● The alternating least squares(ALS) algorithm is
used to iteratively compute a low-rank matrix
factorization.
● The sparse matrix R defines a bipartite graph
connecting each user with the movies that they
rated.
● Vertices are users(rows U) and movies(columns V)
and edges contain the ratings for a user-movie pair.
● GraphLab update function predicts ratings(edge-
values).

•ALS rotates between fixing one of the unknowns ui or vj. When one is fixed the other
can be computed by solving the least-squares problem. The ALS algorithm is as:

•R = {rij }nu ×nv is user-movie matrix where each item Rij represents the rating score of
item
j by user i where ri,j =< ui,vj > ∀i,j.
• U represent the user feature matrix and V represent the movie feature matrix.
• Dimension of the feature space(d) is a system parameter that is determined by a hold-
out dataset or cross- validation.
• The low rank approximation problem is thus formulated as follows to learn the factor
vectors (ui,vj):
Where pi,j =<ui,vj > is the predicted rating ,λ is the regularization coefficient and K is the
Set of known ratings from the Sparse matrix R.

Netflix Scaling with Intensity
•Plotted is the speedup achieved for
varying values of dimensionality(d).
•Extrapolating to obtain the
theoretically optimal runtime, the
estimated overhead of Distributed
GraphLab at 64 machines is 12x for
d=5 and 4.9x for d= 100.

Netflix Comparisons
• GraphLab implementation was
compared against Hadoop and MPI
using between 4 to 64 machines.
•GraphLab performs between 40-60
times faster than Hadoop.
•It also slightly outperformed the
optimized MPI implementation.

Video Co-segmentation(Coseg)
•Video co-segmentation automatically
identifies and clusters spatio-temporal
segments of video that share similar texture
and color characteristics.
•Frames of high-resolution video are pre-
processed by coarsening each frame to a
regular grid of rectangular super-pixels.
•The CoSeg algorithm predicts the best label
(e.g., sky, building, grass etc.) for each super
pixel using Gaussian Mixture Model (GMM)
in conjunction with Loopy Belief Propagation
(LBP).

• The two algorithms are combined to form an Expectation-Maximization problem
alternating between LBP to compute the label for each super-pixel given the GMM and
then updating the GMM given the labels from LBP.
• The GraphLab update function executes the LBP local iterative update where updates
expected to change values significantly are prioritized.The BP update function is as:

• The locking engine provides nearly optimal weak scaling: the runtime does not
increase significantly as the size of the graph increases proportionately with
the number of machines.
• It was also observed that increasing the pipeline length increased
performance significantly and compensated for poor partitioning.

Named Entity Recognition(NER)
• Named Entity Recognition is the task of
determining the type (e.g., Person, Place, or
Thing) of a noun-phrase (e.g. Obama, Chicago,
or Car) from its context (e.g. “President..”,
“Lives near..”, or “bought a..”).
• The Data graph for NER is bipartite with one
set of vertices corresponding to noun-phrases
and the other to contexts.
•The CoEM vertex program updates estimated
distribution for a vertex (either noun-phrase or
context) based on the current distribution for
neighboring vertices.

•Below is CoEM algorithm in which adjacent vertices are rescheduled ,if the
type distributions at a vertex changes by more than some predefined threshold.

NER Comparisons
•GraphLab implementation of NER
achieved 20-30x speedup over
Hadoop and was comparable to the
optimized MPI.
•GraphLab scaled poorly achieving
only a 3x improvement using 16x
more machines, majorly due to
large vertex data size, dense
connectivity, and poor partitioning.

Comparison(Netflix/CoSeg/NER)
• Overall network utilization: Netflix and CoSeg have very low bandwidth requirements
while NER appears to saturate when #machines > 24.
• Snapshot overhead: Overhead of performing a complete snapshot of the graph every|V |
updates is highest for CoSeg, when running on a 64 machine cluster.

EC2 Cost Evaluation
• The price-runtime curves(log-log scale)
for GraphLab and Hadoop illustrate the
cost of deploying either system.
• For the Netflix application, GraphLab is
about two orders of magnitude more
cost-effective than Hadoop.

Conclusion
• In the paper we talked about:
• Requirement of MLDM algorithms.
• Graphlab extended to the distributed setting by:
•Relaxing the schedule requirements
•Introducing a new distributed data-graph
•Introducing new execution engines
•Introducing fault tolerance.
•Distributed Graphlab outperforms Hadoop by 20-60x and is competitive with
tailored MPI implementations.

Future Work
• Extending abstraction and runtime to support dynamically evolving graphs and
external storage in graph databases.
• Further research into theory and application of Dynamic asynchronous graph
parallel computation thus helping in defining emerging field of big learning.

References
• Y. Low, J. Gonzalez, A. Kyrola,D. Bickson, C. Guestrin, and J. M. Hellerstein. Graphlab: A
new parallel framework for machine learning. In UAI, pages 340–349, 2010.
• Y. Low, J. Gonzalez, A. Kyrola,D. Bickson, C. Guestrin, and J. M. Hellerstein.Distributed
Graphlab: A Framework for Machine Learning and Data Mining in the Cloud.
• Thesis:Parallel and Distributed Systems for Probabilistic Reasoning
Joseph Gonzalez,CMU-ML-12-111 ,December 21, 2012.
• GraphLab ppt by Yucheng Low, Joseph Gonzalez, Aapo Kyrola,Danny Bickson,Carlos
Guestrin.
•Y. Zhou, D. Wilkinson, R. Schreiber, and R. Pan. Large-scale parallel collaborative filtering
for the netflix prize. In AAIM, pages 337–348, 2008.
• Christopher R. Aberger. Recommender: An Analysis of Collaborative Filtering Techniques.
• GraphLab create : http://graphlab.org.
• CS-425 Distributed Systems ppt: Chandy-Lamport Snapshot Algorithm and Multicast
Communication by Klara Nahrstedt
• CSE 547 Graph Parallel Problems Synchronous vs Asynchronous Computation pdf by Emily Fox.

GraphLab

More Related Content

What's hot (20)

Viewers also liked (12)

Similar to GraphLab (9)

Recently uploaded (20)

GraphLab