Messaging specialist 29West has published results of performance tests on its IPC (Inter-Process Communication) transport, which push the latency of message transfer to well below one microsecond.
29West's IPC relies on shared memory on a single server to transfer data, instead of communicating between servers via a network. The advent of multi-core and multi-socket architectures by Intel and AMD have made the running of several applications on a single server a reality that allows for very low latencies.
In 29West's tests, a latency as low as 660 nanoseconds was recorded, though this was for message sizes of just 16 bytes, which even 29West officials admit is unrealistic in the real financial world. But for more realistic message sizes used for market data applications, latencies were also very low. As an example, a message size of 256 bytes was communicated in just 763 nanoseconds.
Those numbers are for a 1:1 test - with one server process and one receiver process. Throughput, in terms of messages per second, was also measured for 1:1 and 1:11 configurations, and for 256-byte messages, the respective throughput was 1.58 million and 9.66 million messages per second.
The actual server configuration used for the tests was a Cisco Systems Unified Computing System server, which comprised two Intel Xeon 5670 chips, clocking at 2.96GHz. Each chip offers six cores, for a total of 12 (hence the 1 to 11 fan-out test). The server was configured with 96 Gigabytes of RAM, and ran 64-bit Red Hat Enterprise Linux version 5. The full test report is here.
Using shared memory on a single multi-core server is an increasingly popular architecture being pitched by messaging companies as they seek to reduce latency. NYSE Technologies and Tibco Software have also publicly moved in this direction, while executives at Real-Time Innovations (RTI) say it's been a feature of its offering for some time, though not one they have particularly highlighted.