UME Messaging Performance

UME sheds the extra steps and processes typically found among legacy persistent messaging systems, resulting in significantly lower latency and a true application-to-application product.

As shown below, this design provides distinct advantages in performance. First, an increasing message size hardly impacts average latency at all. Second, because the disk writes are not in the data path, the type of disks that persist the data don't impact latency to any great degree, either.

The graph below shows latency vs. message size results for UME tests run on commodity hardware and released in our white paper "29West UME Persistent Performance on Commodity Hardware with Gigabit Ethernet".

UME Latency vs. Msgsize

Again, these results were achieved on commodity commercial-grade hardware, on a 1G Ethernet link. And while these numbers were obtained with solid-state disks, the results with regular spinning disks were only a couple of microseconds worse for one or two data points.

The data was collected using the following hardware for the UME persistent store:

  • Dual AMD Opteron 2354 Barcelona quad-core processors @ 2.2 GHz
  • 4GB RAM
  • Linux RHEL 5

and the following hardware for the source and receiver machines:

  • Intel Core 2 Quad™ Q6600 processor @ 2.4 GHz
  • 2GB RAM
  • Linux RHEL 5

Testing was done using UME 2.1.1. For more details, see the report linked above.