Hi Avery,
Thanks for the response. I reached out to the graph user mailing list because I
am quite interested in helping develop / execute standardized scalability
testing for Giraph, so I'm glad to see that there is interest!
Here's some follow up to some of the points you raised / questions you asked:
Currently, the biggest limitation faced by GoldenOrb is the capacity issue; it
can't handle more than roughly 100,000 vertices per node. This low maximum
vertices per node limitation, coupled with instability issues, obviously
hampered the ability to conduct ideal scalability testing, but even with graphs
totaling 100,000 to 250,000 vertices a clear power law slope can be found
before hitting an inevitable communication bottleneck. This can be seen by
noting that the log-log slopes of the 20k, 50k, and 100k graphs (for SSSP)
remain fairly constant, and negative, as the number of nodes in the cluster
grows, unlike the slopes for 5k, 2k, and 1k graphs which demonstrate a
framework overhead bottleneck, corresponding to the point where the slope
changes from negative to roughly 0 or positive (which appears to happen at
around 1k vertices per node).
On to the second issue you brought up…
Graph problems can be notoriously difficult to implement scalability testing
for precisely for the reasons you brought up. A few things were done to allow
an apples-to-apples comparison with the Pregel results. First, the single
source shortest path algorithm used for testing comes directly from the Pregel
paper. Second, just as in the Pregel tests, binary tree graphs were used to
ensure that each vertex had the same fixed, low order, outdegree. Last, the
tests were repeated using non-binary tree graphs (generated by a python script)
with a non-constant, but low order, average outdegree per vertex (average 10
edges per vertex, then again with graphs averaging 90 edges per vertex), the
results of which were seen to be quite close to the binary tree graph data.
As mentioned in passing, the scalability test results allow for a direct
comparison with the Pregel results, but should also allow for a meaningful
comparison to your scalability results for Giraph precisely because the edges
per vertex have been fixed. While this is not ideal (I would prefer a
standardized set of tests which everybody runs in standardized configurations),
the proposition that the results can be meaningfully compared is backed up by
two points; First, the log-log slope of the data you presented is right in line
with the value reported by Pregel for their SSSP tests, both of which are
realistic values (and show very good parallelization!), meaning that both
algorithms display similar properties for configurations in the regime not
dominated by a framework overhead bottleneck. And second, the GoldenOrb SSSP
results being compared are also from configurations which have reached a steady
power law slope over the range of nodes considered, for runs using the same
algorithm as the Pregel results. These two points, I feel, justify the
comparisons made (though, again, it would be better to have a standardized set
of configurations for testing to facilitate comparing results, even between
algorithms). Since all three sets of scalability tests yield fairly linear
complexity plots (execution time vs. number of vertices in the graph, slide 29
of your talk), it also makes sense to compare weak scaling results, a
proposition supported by the consistency of the observed GoldenOrb weak scaling
results for SSSP across multiple test configurations.
As for the results found in your October 2011 talk, they are impressive and clearly
demonstrate an ability to effectively scale to large graph problems (shown by the weak
scaling slope of ~ 0.01) and to maximize the benefit of throwing additional computational
resources at a known problem (shown by the strong scaling slope of ~ -0.93), so I'm
interested to see the results of the improvements that have been made. I'm a big
proponent of routine scalability testing using a fixed set of configurations as part of
the software testing process, as the comparable results help to quantify
"improvement" as the software is developed further and can often help to
identify unintended side effects of changes / find optimal configurations for various
regimes of problems, and would like to see Giraph succeed, so let me know if there's any
open issues which I might be able to dig into (I'm on the dev mailing list as well,
though haven't posted there).
Thanks,
Jon
On Dec 11, 2011, at 1:02 PM, Avery Ching wrote:
Hi Jon,
-golden...@googlegroups.com (so as to not clog up their mailing list uninvited)
First of all, thank you for sharing this comparison. I would like to note a
few things. The results I posted in October 2011 were actually a bit old (done
in June 2011) and do not have several improvements that reduce memory usage
significantly (i.e. GIRAPH-12 and GIRAPH-91). The number of vertices loadable
per worker is highly dependent on the number of edges per worker, the amount of
available heap memory, number of messages, the balancing of the graph across
the workers, etc. In recent tests at Facebook, I have been able to load over
10 million vertices / worker easily with 20 edges / vertex. I know that you
wrote that the maximum per worker was at least 1.6 million vertices for Giraph,
I just wanted to let folks know that it's in fact much higher. We'll work on
continuing to improve that in the future as today's graph problems are in the
billions of vertices or rather hundreds of billions =).
Also, with respect to scalability, if I'm interpreting these results correctly,
does it mean that GoldenOrb is currently unable to load more than 250k vertices
/ cluster as observed by former Ravel developers? if so, given the small tests
and overhead per superstep, I wouldn't expect the scalability to be much
improved by more workers. Also, the max value and shortest paths algorithms
are highly data dependent to how many messages are passed around per superstep
and perhaps not a fair scaling comparison with Giraph's scalability designed
page rank benchmark test (equal messages per superstep distributed evenly
across vertices). Would be nice to see an apples-to-apples comparison if
someone has the time...=)
Thanks,
Avery
On 12/10/11 3:16 PM, Jon Allen wrote:
Since GoldenOrb was released this past summer, a number of people have asked
questions regarding scalability and performance testing, as well as a
comparison of these results with those of Giraph (
http://incubator.apache.org/giraph/ ), so I went forward with running tests to
help answer some of these questions.
A full report of the scalability testing results, along with methodology
details, relevant information regarding testing and analysis, links to data
points for Pregel and Giraph, scalability testing references, and background
mathematics, can be found here:
http://wwwrel.ph.utexas.edu/Members/jon/golden_orb/
Since this data will also be of interest to the Giraph community (for
methodology, background references, and analysis reasons), I am cross posting
to the Giraph user mailing list.
A synopsis of the scalability results for GoldenOrb, and comparison data points
for Giraph and Google's Pregel framework are provided below.
The setup and execution of GoldenOrb scalability tests were conducted by three
former Ravel (http://www.raveldata.com ) developers, including myself, with
extensive knowledge of the GoldenOrb code base and optimal system
configurations, ensuring the most optimal settings were used for scalability
testing.
RESULTS SUMMARY:
MAX CAPACITY:
Pregel (at least): 166,666,667 vertices per node.
Giraph (at least): 1,666,667 vertices per worker.
GoldenOrb: ~ 100,000 vertices per node, 33,333 vertices per worker.
STRONG SCALING (SSSP):
Note: Optimal parallelization corresponds to the minimum value -1.0. Deviation
from the minimum possible value of -1.0 corresponds to non-optimal
parallelization.
Pregel: -0.924 (1 billion total vertices)
Giraph: -0.934 (250 Million total vertices)
GoldenOrb: -0.031 Average, -0.631 Best (100000 total vertices), 0.020 Worst
(1000 total vertices)
WEAK SCALING (SSSP):
Note: Optimal weak scalability corresponds to the value 0.0. Deviation from the
optimal value of 0.0, corresponds to non-optimal usage of computational
resources as managed by the framework.
Pregel: No Data Available
Giraph: 0.01 (1,666,667 vertices per worker)
GoldenOrb: 0.37 Average, 0.23 Best (500 vertices per node), 0.48 Worst (12500
vertices per node)
I hope this answers some of the many questions which have been posted regarding
scalability and performance. Be sure to check out the full scalability testing
report at http://wwwrel.ph.utexas.edu/Members/jon/golden_orb/ Please let me
know if you have any questions.
Thanks,
Jon
