Skip to content
Browse files
fix typo, center image
  • Loading branch information
zrlkau committed Dec 11, 2017
1 parent de07803 commit f8588e5ba13719f077b194f95b8cbd44194f0e87
Showing 1 changed file with 2 additions and 7 deletions.
@@ -45,7 +45,7 @@ Crail can accelerate various I/O tasks of Spark Applications, namely shuffle, br

The next figure shows a high level view of CoCoA's communication pattern with 2 workers and 1 driver process. Blue arrows represent broadcast communication whereas green arrows represent a reduce operation.

![communication pattern of the cocoa algorithm](/img/blog/crail-machine-learning/cocoa.svg)
<div style="text-align:center"><img src ="/img/blog/crail-machine-learning/cocoa.svg" width="550"/></div>

The first step in accelerating this workload was to analyze where time is being spent. An initial breakdown resulted in this

@@ -134,15 +134,10 @@ deserialized and copied multiple times for every reduce step. Using Crail, it wa
modify this code for zero-copy and single serialization. This reduced the time spent in the reduce
path from 1185 ms to 680 ms (1.74x faster).

Maybe mention the custom serializer somehow, though it didn't make a huge difference, but it allowed
to use Crail's zero-copy serialization.

### Putting it all together
When using all optimizations we were able to reduce the runtime of the Snap.ML machine learning
application from ~2.5s to 1.4s per iteration (1.77x faster).

Final plot with all data

0 comments on commit f8588e5

Please sign in to comment.