Researchers find 10 Gigabit Ethernet as speedy as customized interconnects

A Los Alamos National Laboratory researcher working with fellow academics has shown that 10 Gigabit Ethernet adapter cards can offer nearly the same performance as other specialized high-performance network interconnects.

'Our evaluations show that in most experimental scenarios, 10GigE provides comparable (or better) performance than [InfiniBand] and Myrinet,' according to a research paper co-authored by Wu Feng, working with researchers at Ohio State University.

Feng is a technical staff member and team leader for Los Alamos National Laboratory, though he will leave this position shortly to join the faculty of the Department of Computer Science and to be a founding member of the Center for High-End Computing Systems, both at Virginia Tech.

Feng's findings run contrary to conventional wisdom in the high-performance computing community that high-performance interconnects such as InfiniBand and Myrinet offer significant superior performance over Gigabit Ethernet and the newly emerging 10 Gigabit Ethernet.

Gigabit Ethernet is already the most widely used interconnect in high-performance computing environments. In the November iteration of the Top 500 list of fastest supercomputers, compiled by the universities of Tennessee and Mannheim (Germany) and National Energy Research Scientific Computing Center, almost 50 percent of these computers used Gigabit Ethernet interconnects. The second most popular interconnect was Myrinet, supplied by Myricom Inc. of Arcadia, Calif.

These organizations use Gigabit Ethernet because it is less expensive to deploy, according to the researchers. Traditionally, it has been assumed that systems architects sacrifice performance for the more attractive price tag, however. In particular, the interconnect has dramatically higher latency'to the speed in which an interconnect can relay a packet of data. Yet, specialized interconnect cards can run as much as $800 each, while Gigabit Ethernet can run for as low as a few dollars each, thanks to widespread adoption. The 10GigE cards are expected to follow suit.

To boost 10GigE performance, a number of vendors have introduced 10GigE cards that come with a processor known as a TCP Offload Engine, which handles some of the work processing TCP packet headers'a job normally reserved for the central processor.

Vendors claim that TOE-based cards perform on par with interconnects dedicated to HPC environments, such as InfiniBand and Myrinet. Testing such claims, however, has been difficult, given the differences between the technologies.

Feng's research team developed a test to compare the three networking technologies. They recorded the latencies of Myrinet, InfiniBand and 10GigE adapters at three uniform points'at the native socket layer, at the Message Passing Interface layer atop the socket and at the application layer. This final test measured how quickly the data was transferred to and from a biomedical image visualization tool, an oil reservoir simulator and a cluster management tool.

The group had found that for an event-based model, 10GigE surprisingly held its own. Myrinet had achieved a latency of 11.33 microseconds, followed by 17.7 microseconds by 10GigE and 24.4 microseconds of an InfiniBand interconnect. Moreover, the Myrinet latency increased notably as the message sizes got larger.

Feng admitted that the type of notification mechanism used in these tests was not one usually used by Myrinet and InfiniBand. This 'event'-based mechanism, used in 10GigE, is based on the adapter alerting the operating system socket when a new packet has arrived'an approach that is slower but uses a processor more efficiently.

Both InfiniBand and Myrinet use a 'polling' mechanism, in which the socket continuously monitors a space in memory to check if a data transmission has been completed. In this mode, the team found that Myrinet had a 4.64 microsecond latency and InfiniBand at 8.25 microsecond latency'both beating 10GigE by a wide margin.

When it comes to bandwidth, 10GigE championed, at least in these tests. Bandwidth is the amount of data an interconnect can handle in any one given period of time. In these tests, 10GigE channeled in one direction 6.4 Gbps, while InfiniBand did 5.4 Gbps and Myrinet did 3.9 Gbps. Unidirectional traffic showed similar numbers.

Feng admits that the since these results were published, InfiniBand and Myrinet providers have lowered latency times, although he notes that work is being done to streamline 10GigE through a specification called iWarp, being developed by the Remote Direct Memory Access Consortium. Also, Ethernet switch makers are starting to produce products that do virtual cut-through routing favored by HPC switch makers rather than the slower, traditional, store-and-forward mechanisms now being used. This work should also lower network latency, he said.

'Ethernet is a million-pound gorilla. Given the way that we are being driven to cheaper commodity parts, I [suspect] that the companies that adapt 10 Gigabit Ethernet are the ones that survive,' Feng said.