STAC – aka the Securities Technology Analysis Center – has benchmarked IBM’s proprietary Platform Symphony implementation of Hadoop MapReduce, versus the standard open source offering, to compare their respective performance. On average, IBM’s implementation performed jobs 7.3 times faster than the standard, reducing total processing time by a factor of six.
Better known for its benchmarking of low-latency trading platforms, STAC leveraged the Statistical Workload Injector for MapReduce (SWIM), developed by the University of California at Berkeley. SWIM provides a large set of diverse MapReduce jobs based on production Hadoop traces obtained from Facebook, along with information to enable characterisation of each job. STAC says it undertook the benchmarking because many financial markets firms are deploying Hadoop.
The hardware environment for the testbed consisted of 17 IBM compute servers and one master server communicating over gigabit Ethernet. STAC compared Hadoop version 1.0.1 to Symphony version 5.2. Both systems ran Red Hat Linux and used largely default configurations.
IBM attributes the superior performance of its offering in part to its scheduling speed. IBM’s Hadoop is API-compatible with the open source offering but runs on the Symphony grid middleware that became IBM’s with its aquisition of Platform Computing, which closed in January of this year.
For more information on STAC’s IBM Hadoop benchmark, see here.