Since 1986 - Covering the Fastest Computers in the World and the People Who Run Them

Language Flags
March 25, 2010

A Couple of New InfiniBand Songs

Michael Feldman

Apropos of nothing in particular, it occurred to me that “InfiniBand” would actually be a good name for a rock band. OK, to be honest, the topic is apropos of something. I’ve been spilling a lot of digital ink lately on the new high-end CPU and GPU news this spring, so I’ve let a few InfiniBand stories slip by. With that in mind, I thought this would be a good opportunity to recap a couple of the latest happenings in the IB universe.

This past Tuesday, QLogic announced a QDR InfiniBand pass-through module for Dell PowerEdge blade systems. The new QLogic 12005-PT16 is a single-slot module that fits neatly into a PowerEdge M1000e chassis. The module supports up to 16 QDR ports that directly hook into an InfiniBand mezzanine card. Conveniently, QLogic is also introducing a M1000e mezz card, the QME7342, that fulfills this role. The mezz card connects the blade servers to a standard external InfiniBand switch (QLogic or otherwise).

So what’s the big fuss? If you missed the pass-through module lecture in your Networking 101 class, here’s the Cliff Notes version: Pass-through hardware is an alternative to a traditional internal switch found in most blade systems. Unlike a switch, which directs message traffic to specific network addresses, a pass-through is a much simpler solution that just provides a point-to-point connection from the host out to the external fabric. The general idea of a pass-through is to make the switches independent of the compute infrastructure, just as they are in a rack server setup.

The limitation is that blade architectures tend to be very proprietary, and not all systems are amenable to a pass-through solution. It depends on how the OEM has designed the enclosure, the ability to support the additional cabling required by a pass-through solution, and also if that vendor is interested in integrating and selling switch hardware as part of its solution. In the case of the M1000e, Dell does offer a Mellanox M3601Q switch for InfiniBand connectivity. But the M-Blade design can accommodate a pass-through as well and, according to QLogic, has some big advantages.

One of those advantages is that the QLogic module only occupies a single slot in the blade chassis, instead of two for the Mellanox switch. As a result, there is an extra I/O slot in the blade chassis that can be used for a 10 GbE hookup. This is especially important if the blade needs to access Ethernet storage or I/O out on the network.

Also, because there’s no switch silicon and associated hardware in the pass-through, it uses much less power. The QLogic 12005-PT16 draws just 40 watts compared to 80 watts for the Mellanox switch. Overall, QLogic is claiming a 45 percent power savings for a M1000e enclosure using its pass-through module and new mezz card. This is a big deal, since one of the knocks on blade systems is that they tend to run hot as a result of the dense designs.

Finally, since the pass-through module doesn’t do any switching, it eliminates internal switch hops, which cuts down on communication latency. As long as you don’t have external edge switches to deal with, you can get away with eliminating a whole interconnect layer and connect directly to the core datacenter switch. QLogic reports an almost non-existent 400 picosecond point-to-point latency for the pass-through module itself.

In talking to Steve Zivanic, QLogic’s director of corporate communication, this latest set of products is part of a broader strategy to expand their InfiniBand penetration via their OEM partners. QLogic IB gear is now being stocked by some of the biggest HPC cluster makers, including IBM, HP, Dell, and SGI. According to Zivanic, they’re ramping up the sales team and hiring HPC specialists to further increase their competitiveness. “You’re going to see us getting a lot more aggressive this year with InfiniBand,” he told me.

The other InfiniBand story that slipped by a couple of weeks ago was Voltaire’s announcement of its Fabric Collective Accelerator (FCA) technology to accelerate MPI operations on Voltaire switches. In a nutshell, the FCA software uses CPUs inside the switches to offload MPI collective operations. Voltaire claims that the initial version is able to accelerate “high performance computing applications such as reservoir modeling, fluid dynamics, crash analysis and others by up to a factor of ten.” And from the application’s point of view, the offloaded computations are transparent.

Potentially this could greatly lengthen the longevity of MPI codes. One of the weaknesses of the distributed MPI style of computing is that collective operations have to be globally synchronized, which makes these applications harder to scale. And the routines themselves are recursive, so tend to be extremely compute-intensive. The result is that for some codes up to 90 percent of the host’s CPU time is spent inside these collective routines.

The brute force solution has been to make the switch and host adapter hardware faster to compensate for the larger fabrics and cluster sizes, and make the cluster nodes more computationally dense to speed host-side processing. Mellanox has incorporated MPI offload technology inside a host adapter, but by adding MPI smarts into the switch itself, Voltaire says message processing can optimized much more efficiently.

According to Asaf Somekh, Voltaire’s VP of marketing, the FCA technology not only relieves the host of a computational burden, it also is in a position to cut down on the volume of message traffic being sent across the fabric, thus reducing communication bottlenecks. The idea is to magnify the computational performance of the servers, turning them into a sort of super-cluster.

That would be welcome news indeed for HPC customers who would like to be able to use clusters for capability-level computing, rather than having to turn to the more custom-built (and expensive) solutions like the IBM Blue Gene or the Cray XT. In fact, Somekh believes FCA will be a key technology to propel commodity clusters into the capability supercomputer market. “In my mind, this is one of the Holy Grails of HPC,” he says.

OK. But it deserves a bit stronger branding than Fabric Collective Accelerator. How about “InfiniBand Overdrive”? Hey, now that really would be a great name for a rock band.