STAC Report: GemCache Data Transactions
Two-server stack shows latency <1 ms at 6000 orders per sec, with data replication.
Monday, 14 January 2008 - STAC® has just released its first public STAC Report of a benchmark using a trade-execution workflow. We tested GemCache Data Transactions, an integrated solution from GemStone Systems, IBM, and Intel configured to handle high-speed trade order messages. The focus of the benchmark was to measure latency at different order rates for a system that was configured for zero data loss. The system consisted of GemFire® Enterprise 5.0 software, Red Hat Enterprise Linux 4.4, IBM BladeCenter® HS21 servers with Intel® Xeon® Quad Core 5355 processors, and SDP over Cisco SFS InfiniBand.
We benchmarked GemCache using STAC Tools, which generated order traffic and measured the system’s latency and throughput. The tests were designed to answer three primary questions:
* How fast is a trip through the GemCache solution?
* Does the solution really provide zero data loss?
* What sorts of resources are required to reach 15,000 orders per second while maintaining a mean latency of approximately 1 millisecond?
GemCache Data Transactions performed well:
* Mean latency less than 1 millisecond (987 µsec) at 6000 orders per second, with full data replication using a two-server GemCache configuration.
* Mean latency of just 519 microseconds at 6000 orders per second with a single GemCache server without replication
* Three active/active two-server clusters can support 18,000 orders per second with approximately 1 millisecond (1017 µsec) of end-to-end latency using six IBM Server blades.
* No data loss was observed during any test, including a two-server failover test.
Report download available here to registered users at no cost.
If you’re interested in performance measurement of trade execution solutions, consider joining the STAC Benchmark Council. We are currently finalizing the date for the first meeting of the STAC-E Benchmark working group, which is focused on defining benchmark specifications for execution-related workflows. For more information on STAC Benchmarks or the Council, see www.STACresearch.com/council.