Engineering Distributed Graph
Algorithms in PGAS languages
Guojing Cong, IBM research
Joint work with George Almasi
and Vijay Saraswat
Programming language from the perspective
of a not-so-distant admirer
Mapping graph algorithms onto distributed
memory machines has been a challenge
•
•
•
•
Efficient mapping PRAM algorithm onto SMPs is hard
Mapping onto a cluster of SMPs is even harder
Optimizations are available and shown to improve performance
Can these be somehow automated with help from the language
design, compiler and runtime development?
• Expectations of the languages
– Expressiveness
• SPMD, task parallelism (spawn/async), pipeline, future, virtual sharedmemory abstraction, work-stealing, data distribution, …
• Ease of programming
– Efficiency
• Mapping high level constructs to run fast on the target machine
–
–
–
–
SMP
Multi-core, multi-threaded
MPP
Heterogeneous with accelerators
• Leverage for tuning
A case study with connected components on
a cluster of SMPs with UPC
• A connected component of an undirected graph G=(V,E), |V|=n,
|E|=m, is a maximal connected subgraph
– Connected components algorithm find all such components in G
• Sequential algorithms
– Breadth-first traversal (BFS)
– Depth-first traversal (DFS)
• One parallel algorithm -- Shiloach-Vishkin algorithm (SV82)
– Edge list as input
– Adopts the graft and shortcut approach
• Start with n isolated vertices.
• Graft vertex v to a neighbor u with (u < v)
• Shortcut the connected components into super-vertices and continue on the
reduced graph
Example: SV
4
2
4
2
1,4
1st iter.
2nd iter.
1
3
Input graph
1
1
1
2
2,3
3
graft
shortcut
2
Simple? Yes, performs poorly
Random Graph, 1M vertices, 10M edges
Random Graph (1M vertices, 20 M edges)
300
100
Bor-AL
Bor-EL
Prim
SV
Time (seconds)
Execution Time (400M cycles)
250
10
200
150
100
50
1
0
2
4
6
Number of Processors
8
10
12
2
4
6
8
Number of Processors
Sun enterprise E4500
• Memory-intensive, irregular accesses, poor temporal
locality
10
Typical behavior of graph algorithms
• CPI construction
• BC – betweeness centrality
• BiCC – Biconnected
components
• MST – Minimum spanning
tree
• LRU stack distance plot
On distributed-memory machines
• Random access and indirection make it hard to
– implement, e.g, no fast MPI implementation
– Optimize, i.e., random access creates problems for both
communication and cache performance
• The partitioned global address space (PGAS) paradigm
– presents a shared-memory abstraction to the programmer for
distributed-memory machines. receives a fair amount of attention
recently.
– allows the programmer to control the data layout and work
assignment
– improve ease of programming, and also give the programmer
leverage to tune for high performance
Implementation in UPC is
straightforward
UPC implementation
Pthread implementation
Performance is miserable
Communication efficient algorithms
•
Proposed to address the “bottleneck of processor-to-processor
communication”
– Goodrich [96] presented a communication-efficient sorting algorithmon weakCREWBSP that runs in O(log n/ log(h + 1)) communication rounds and O((n log
n)/p) local computation time, for h = Θ(n/p)
– Adler et. al. [98] presented a communication-optimal MST algorithm
– Dehne et al. [02] designed an efficient list ranking algorithm for coarse-grained
multicomputers (CGM) and BSP that takes O(log p) communication rounds with
O(n/p) local computation
•
Common approach
– simulating several (e.g., O(log p) or O(log log p) ) steps of the PRAM algorithms
to reduce the input size so that it fits in the memory of a single node
– A “sequential” algorithm is then invoked to process the reduced input of size
O(n/p)
– finally the result is broadcast to all processors for computing the final solution
•
Question
– How well do communication efficient algorithms work on practice?
– How fast can optimized shared-memory based algorithms run? Cache
performace vs. communication performance
– Can these optimizations be automated through necessary language/compiler
support
Locality-central optimization
• Improve locality behavior of the algorithm
– The key performance issues are communication and
cache performance
– Determined by locality
• Many prior cache-friendly results, but no tangible
practical evidence
– Fine-grain parallelism makes it hard to optimize for
temporal locality
– Focus on spatial locality
• To take advantage of large cache lines, hardware
prefetching, software prefetching
Scheduling of the memory accesses in a
parallel loop
Typical loop in CC
Generic loop
An example
Mapping to the distributed
environments
• All remote accesses are
consecutive in our
scheduling
• If the runtime provides
remote prefetching or
coalescing, then
communication efficiency
can be improved
• If not, coalescing can be
easily done at the
program level as shown
on right
Performance improvement due to
communication efficency
Applying the approach to single-node
for cache-friendly design
CC
0.75
100M, 400M
100M, 1G
200M, 800M
0.70
normalized execution time
• Apply as many levels of
recursions as necessary
• Simulate the recursions
with virtual threads
• Assuming a large-enough,
one level, fully associative
cache
0.65
0.60
0.55
0.50
0.45
Original execution time
0.40
1
Optimized execution time
2
4
6
8
10
W
12
14
16
18
20
Graph-specific optimization
• Compact edge list
– the size of the list determines the number of elements
to request from remote nodes
– edges within components no longer contribute to the
merging of connected components, and can be
filtered out
• Avoid communication hotspot
– Grafting in CC shoots a pointer from a vertex with
larger numbering to one with smaller numbering.
– Thread thr0 owns vertex 0, and may quickly become
a communication hotspot
– Avoid querying thr0 about D[0]
UPC specific optimization
• Avoid runtime cost on local data
– After optimization, all direct access to the shared arrays are local
– Yet the compiler is not able to recognize
– With UPC, we use private pointer arithmetics for
• Avoid intrinsics
– It is costly to invoke compiler intrinsics to determine the target
thread id
– Computing target thread ids is done for every iteration.
– we compute these ids directly instead of invoking the intrinsics.
– Noticing that the target ids do not change across iteration, we
compute them once and store them in a global buffer.
Performance Results
Random Graph, 100M vertices, 400M edges
Random Graph, 100M vertices, 1G edges
1000
1000
Optimized
SMP
BFS
Time (seconds)
Time (seconds)
Optimized
SMP
BFS
100
10
100
10
16
32
64
128
256
16
32
# Threads
Hybrid Graph, 100M vertices, 400M edges
128
256
Random Graph, 100M vertices, 400M edges
100
120
Comm
Sort
Copy
Irregular
Work
Setup
60
40
20
Comm
Sort
Copy
Irregular
Work
Setup
100
Time (seconds)
80
Time (seconds)
64
# Threads
80
60
40
20
0
0
base
compact
offload
circular
Implementations
localcpy
id
base
compact
offload
circular
Implementations
localcpy
id
So, how helpful is UPC
• Straightforward mapping of shared-memory
algorithm is easy
– quick prototyping
– Quick profiling
– Incremental optimization (10 versions for CC)
• All other optimizations are manual
• Many of them can be automated, though
• UPC is not flexible enough to expose the
hierarchy of nodes and processors to the
programmer
Conclusion and future work
• We show that with appropriate optimizations, shared-memory graph
algorithms can be mapped to the PGAS environment with high
performance.
• On inputs that fit in the main memory on one node, our
implementation achieves good speedups over the best SMP
implementation and the best sequential implementation.
• Our results suggest that effective use of processors and caches can
bring better performance than simply reducing the communication
rounds
• Automating these optimizations is our future work
Descargar

Engineering Distributed Graph Algorithms in PGAS …