Tervela Breaks 10 Million Messages per Second Barrier in First Real World Test Environment
Tervela, provider of high-performance messaging systems, says its Tervela Message Network, officially launched in September 2007 (Low-Latency.com, September 19 2007), can process more than 10 million messages per second in a real-world test environment. Tervela’s fault-tolerant message switches can route US market data – including all equities, derivatives, commodities and FX instruments – and send it out to more than 1500 subscribers.
The mean baseline roundtrip* message latency was 58 microseconds with a standard deviation of 8 microseconds. The mean latency numbers increased by only 18 microseconds during peak load conditions – even during sensitive phases such as market open and close – which are subject to radical changes in volatility. The test conditions were constructed using parameters that best matched the operations of most large financial services firms including Gigabit Ethernet, 150 byte messages and UDP point-to-point transport. The tests were performed at Advanced Micro Devices’ (AMD’s) Sunnyvale, California headquarters.
* Each publishing server also runs a subscribing application, which receives certain traffic in order to measure latency. Running publishers and subscribers on the same server means that the timestamp clock is the same, and avoids the need to reconcile clocks across multiple servers.
Technorati Tags: Tervela, high-performance messaging systems, Tervela Message Network, fault tolerant message switches, mean latency, Gigabit Ethernet, Advanced Micro Devices, AMD
March 5th, 2008 at 7:41 am
10million messages a second, across 100 servers… that’s 100,000 messages per server. Let’s assume they are basic 4 core boxes. That’s 25,000 messages per core. Even if youdouble that- is this impressive? RV can do that handily.
Latency figures are low, certainly.
It seems like this “switch” is a choke point though, no? I don’t get it.
March 10th, 2008 at 1:34 pm
Are the 100 servers for the messaging, or to simulate the 1500 subscribers?
April 10th, 2008 at 8:54 am
I’ve spent some time with Tervela understanding the details of this test. The 100 servers are simulating the publishers and subscribers - there are 15 applications per server, 1 publishing and the others subscribing.
As for your criticism, “Skeptic”, you’re failing to understand the point of the test. The test is not to demonstrate application capacity, it’s to demonstrate switch performance under intense messaging conditions. In our tests of Tervela, we’ve seen their API out-perform any other messaging API, including RV and 29West (I’m not including MAMA or PAPA because they’re not really competitive as far as messaging goes) from both a throughput and latency perspective, which we found incredibly impressive considering that Tervela puts an extra hop between API clients.