It was a bit of a surprise when QLogic beat out Mellanox as the interconnect vendor on the National Nuclear Security Administration’s (NNSA’s) Tri-Lab Linux Capacity Cluster 2 contract in June. Not only was Mellanox the incumbent on the original Tri-Lab contract, but it is widely considered to have the more complete solution set for InfiniBand. Nevertheless, QLogic managed to win the day, and did so with somewhat unconventional technologies.
One of these is QLogic’s TrueScale InfiniBand architecture. TrueScale uses an on-load approach to networking in which the lion’s share of packet processing is passed off to the CPUs on the servers. That allows host channel adapters (HCAs) based on TrueScale chips to be much simpler in design than those used to offload those functions (in particular Mellanox ConnectX-based adapters), but at the cost of using CPU resources to do network tasks like packet processing.
That’s why offloading has been the traditional answer for computationally-burdened HPC systems, not just for lower-level packet manipulation, but for MPI processing as well. And it makes perfect sense. The less communication processing the CPUs have to do, the more time they can spend on the application.
But it doesn’t always work out that way in the real world. Especially for certain types of codes where the bottleneck is communication, rather than computation, being able to tap into host CPUs can be an advantage. This is especially true in modern-day clusters, which are filled with core-rich CPUs, not all of which can be fully utilized 100 percent of the time. In these situations, on-loading can exploit essentially free cycles and in a manner that scales naturally with the size of the cluster.
But even where the application is more computationally intensive, QLogic maintains that its on-load approach will still outrun Mellanox’s offloading approach. They attribute that to the other critical piece of their InfiniBand technology: Performance Scaled Messaging (PSM). PSM is QLogic’s communication library that it touts as their lightweight answer to InfiniBand Verbs. The latter was defined by the original InfiniBand spec designers to provide a general-purpose communication API that assumed RDMA and some sort of offloading in the network adapter.
QLogic came up with PSM as a leaner, meaner interface designed explicitly for high performance computing. And now that PSM has been turned over as open source and incorporated into the OpenFabrics Enterprise Distribution (OFED), the software can now be embraced by the wider HPC community. Like Verbs, PSM is supported in all major MPI implementations.
According to Joseph Yaworski, director of HPC Product and Solution Marketing at QLogic, PSM is what makes their InfiniBand offering so efficient for HPC environments. Both PSM and Verbs run on the server CPUs, but unlike Verbs, which was originally designed for handling of I/O requests in a datacenter environment (and later modified to support message passing when HPC became the primary user of InfiniBand), PSM was purpose-built for MPI from the start.
The difference is the nature of the communication for the two application areas. While I/O usually entails relatively large blocks of data to be sent across a limited number of nodes, MPI communication often requires tens of millions of relatively small messages to be passed between hundreds or even thousands of CPU cores.
“Verbs, due to its poor semantic match between MPI’s message passing requirements and the structure of the Verbs implementation, means that a heavy weight protocol must be traversed to handle each message,” says Yaworski. “This approach puts a significant burden on the host CPU and severely limits network performance, especially as a cluster is scaled.
QLogic points to a couple of ANSYS FLUENT benchmarks to show its InfiniBand performance on these common CFD codes. The tests were run on a 384-core server cluster, made up of 32 computational nodes and one NFS server node. Each server consisted of dual quad-core Intel Xeon 5670 “Westmere” 2.93GHz processors and 24GB of memory. Platform MPI was used with the MPI stats option turned on to collect the statistics for communications and CPU utilization. According to Yaworski, the same object code was used for the application for both on-loading and offloading runs.
The first test was the Eddy 417K cell model, which is relatively light on the computation side, but heavy on the communications. For this application, QLogic says on-loading with PSM delivers 366 percent more application performance than offloading with Verbs, claiming the difference is the more efficient use of the CPUs. With this model, just 76 percent of the CPU cycles were used for communication with on-loading/PSM versus 95 percent for offloading/Verbs.
The second FLUENT test case is the Truck 111M cell model, which is much more computationally intensive. In this case, the QLogic solution runs just 20 percent faster, since the overall communication burden is much less, although still taking up 53 percent of the CPU for on-loading with PSM and 61 percent for offloading with Verbs.
As one might suspect, Mellanox is having none of this. According to Gilad Shainer, senior director of HPC and Technical Computing at Mellanox, the offloading critique is unfounded, and benchmark tests such as the ones QLogic touts can be easily manipulated for the benefit particular outcomes. From his perspective, QLogic’s positioning of their InfiniBand on-load technology is a marketing ploy to make up for the lack of sophistication in the TrueScale silicon.
Shainer maintains that the rationale for offloading is straightforward: to be able to use system resources for what they do best, in this case, CPUs for computation and HCAs for network processing. According to him, that’s why most adapters use some form of offloading today, whether to support InfiniBand and MPI communication, Fibre Channel over Ethernet, TCP offload, or what have you.
On-loading also makes RDMA (Remote Direct Memory Access) impossible, which means data must be buffered by the CPU in certain situations, instead of being directly mapped by the HCA. In those cases, data transfer latencies are much higher — up to 7 times higher according to Mellanox — and throughput is lower.
This is especially true when InfiniBand is used to connect storage. Shainer says for file system applications like Lustre and GPFS, you can lose up to half the I/O bandwidth without RDMA (Yaworski concedes that Mellanox is currently better for InfiniBand-based storage but says QLogic is within “spitting distance” of its competitor on I/O performance.) Shainer also says RDMA gives Mellanox’s GPUDirect implementation a decided performance advantage, a claim disputed by QLogic.
On the other hand, says Shainer, just because the offload capability is on-chip, there is no requirement to use it. Mellanox supports network transport and MPI offload capabilities, but the user is able to switch those features on and off if so desired. In that sense, he points out, offloading is really a superset of on-loading.
Nevertheless, recent experience on some large clusters at Lawrence Livermore National Lab (LLNL) appear to back QLogic’s claims of scalability and performance, at least on some of the lab’s simulation codes. On Sierra, a 1,944-node HPC cluster at LLNL connected with QLogic InfiniBand adapters and switches, a multiphysics code was able to achieve 1 to 2 us of MPI latency across 24,000 cores and attain 27 to 30 million messages per second. Matt Leininger, deputy for advanced technology projects at LLNL, said that Sierra demonstrated better scaling than any of their other clusters, not to mention their older Blue Gene and Cray XT supercomputers. Leininger attributed the superior performance to the QLogic network.
At LLNL, QLogic InfiniBand now connects more than 4,000 nodes spread across Sierra and three smaller HPC clusters. The lab’s positive experience with the technology was almost certainly a factor that led the NNSA to select QLogic QDR InfiniBand over the Mellanox offering on the second Tri-Labs contract announced. With the win, QLogic gear will now be firmly entrenched at Sandia National Laboratories and Los Alamos National Laboratory, the other two labs in the Tri-Labs complex.
While Mellanox will continue to be a market leader in InfiniBand for the foreseeable future, QLogic may have found a technological strategy that enables it to expand its market share. Continuing to exploit that strategy is going to be tough going, given its competitor’s dominance in the InfiniBand market. But in Mellanox’s more all-encompassing RDMA offerings and QLogic’s more bare-bones HPC approach, the market may have found the differentiation needed to keep both InfiniBand product sets viable.