Is 10 Gigabit Ethernet Ready for HPC?
Despite the still-modest showing of 10 Gigabit Ethernet (10GbE) technology in high performance computing deployments, vendors at SC10 were showcasing a wide array of performance-laden Ethernet products. In addition to stalwart HPC network vendors like Arista, BLADE Network Technologies (now part of IBM), Myricom, Voltaire and Mellanox, plying their 10GbE wares in the exhibition hall, we also saw Cisco, Brocade, Fulcrum, Solarflare, and ADVA Optical Networking trying to woo the HPC faithful.
IT Brand Pulse Labs analyst Tim Dales takes a look at the prospects for 10GbE in high performance computing, the migration pattern from GbE to 10GbE, and some application areas that seem especially suitable for the technology.
There are two key elements in 10 Gigabit Ethernet that will help drive increased adoption in the high performance computing (HPC) market: throughput (Mbps) and low-latency. Until now the networking I/O provided for HPC has been dominated by 1 Gigabit Ethernet (GbE) and InfiniBand. According to the HPC Advisory Council, in 2009 the HPC market interconnect breakdown was 43 percent GbE, 40 percent InfiniBand and 17 percent other. However, this is all about to change as 10GbE, low-latency adapters emerge in the market and become proven in HPC applications.
1Gbit to 10Gbit Migration
An enormous number of legacy HPC systems use 1Gbit today to access storage (NFS) and for inter-processor communication (IPC) between nodes. Upgrades to 10Gbit in HPC implementations can net 10x throughput improvements with a small investment in adapters and switches. B simply upgrading the network I/O, a large compute job in an HPC cluster would take a fraction of the time to complete and system managers can get more work done in less time. The question for existing HPC systems managers is whether the investment in improved network I/O will yield the necessary return to justify the upgrade. Currently, the price delta between 1Gbit and 10Gbit sever adapters is 4x for 10x of throughput improvement, which results in a lower 10Gbit normalized price.
As new HPC projects are funded and launched in the private and public sectors, designers will take a hard look at 10Gbit due to its cross-market appeal and improved performance but need to be convinced that the second element, low-latency capability, is available.
10Gbit Low-Latency Ethernet Applications
The brightest spot in 10Gbit low-latency applications is financial services and a subset called High Frequency Trading (HFT). HFT is a multi-billion dollar business niche that relies on low-latency market feeds from stock exchanges, a cluster of high-powered processors to run proprietary algorithms on the data, and another low-latency TCP connection back to the market to execute split-second trades. The trading firm that gets their trades in first…wins! With millions of dollars riding on daily trades with lightning fast executions you can see how low-latency 10Gbit is a necessity for this business model.
For clarification, low-latency Ethernet is described by a ½ round trip time (RTT) and is the length of time it takes for a signal to be transmitted plus the length of time it takes for the acknowledgment of that signal to be received. That sum is divided by 2 for the average one-way point-to-point latency. For Ethernet this is measured in microseconds (usec.) To give you a reference point, current 1Gbit Ethernet has a typical latency of approximately 20 usec. In the HFT application mentioned, latency was 4-5 usec, a number that I believe will prove acceptable in many HPC applications.
Consider that 10Gbit low-latency (4-5 usec.) financial application we just spoke of and see how it can be applied on a broader sense to HPC applications that have different characteristics, but can still benefit from increased throughput and low latency. For example, a seismic data processor will acquire terabytes of seismic data from the earth, transfer the data to a multi-node cluster and process the data. With 10Gbit low-latency Ethernet in the cluster, the IPC latency between compute nodes in the cluster is reduced which reduces the compute-time for seismic data processing jobs, that could be on the order of hours or even days of time savings. In addition, the 10x throughput improvement results in lower data transfer time from storage to compute cluster. The result is that seismic data reports are delivered quicker, customers are happier, and revenue comes sooner.
The same low-latency and high throughput features apply to other HPC applications, such as plasma physics simulations, life science modeling, and other clustered applications where huge reductions in compute run-times can be realized just by changing the network I/O.
It is well known in the HPC community that low-latency, high-bandwidth systems are critical to success. 1Gbit Ethernet has the mass-market appeal, and comprises a majority of HPC systems, but is not yet seen as a performance leader. I believe the new capabilities offered with 10Gbit: 10x higher data rate, and 5x lower latency will make it successful for HPC systems.
Buyer Beware! Not all 10Gbit server adapters provide very low-latency. Make sure you ask the vendor for the latency specs and any benchmarks they have performed to substantiate their low-latency claims.
About the Author
Tim Dales has over 25 years experience in product management and technical sales of hardware, software and services at BakBone Software, Incentra Solutions, MTI, Emulex, MAI Systems and AT&T. Tim runs IT Brand Pulse Labs which specializes in providing independent, third party testing and lab reports about Unified Networking products.