Details
-
Improvement
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
-
None
-
None
Description
Having a cache on mappers and reducers could be very useful for some use cases, including but not limited to:
1. Iterative Map Reduce Programs: Some machine learning algorithms frequently need access to invariant data (see Mahout) over each iteration of MapReduce until convergence. A cache on such nodes could allow easy access to the hotset of data without going all the way to the distributed cache. This optimization has been described by Jimmy Lin et. al in the paper "Low-Latency, High-Throughput Access to Static Global Resources within the Hadoop Framework" (http://hcil2.cs.umd.edu/trs/2009-01/2009-01.pdf)
2. Storing of intermediate map outputs in memory to reduce shuffling time. This optimization has been discussed at length in Haloop (http://www.ics.uci.edu/~yingyib/papers/HaLoop_camera_ready.pdf), and by Shubin Zhang in "Accelerating MapReduce with Distributed Memory Cache" presented at ICPADS 2009.
There are some other scenarios as well where having a cache could come in handy.
JSR 107 aims to standardize caching interfaces for Java Application and popular caching solutions such as Ehcache and Memcached have JSR 107 wrapper. Hence, tt will be nice to have some sort of pluggable support for JSR 107 compliant caches on Hadoop.