October 06, 2006
InfiniBand and iWARP have industry heavyweights behind them, to be sure. But this kind of smug satisfaction has only led to a case of the Emperor's New Clothes in which no one is willing to admit to the inadequacy of RDMA in general and VIA in particular.
Kernel Bypass, Zero Copy, and Asynchronous Communication
Networks are a shared resource. Traditional networks such as Ethernet require that the resource be protected by the kernel, which presents a tremendous performance bottleneck when latency is an issue. Furthermore, data is copied to and from pre-allocated buffers, which can hurt bandwidth for large messages.
Many of today's high-performance networks from vendors such as Myricom and Quadrics handle the protection across process boundaries directly through the network interface card (NIC). This setup bypasses the kernel and lets communication occur at the user level, thereby removing the bottleneck of mode switching.
Most modern high-performance networks also have direct memory access (DMA) in which the NIC accesses main memory directly while the CPU is free to perform other tasks. DMA not only eliminates copying, but also permits communication to overlap with computation. This facet is similar to prefetching in the cache as it reduces the effective latency. Taking advantage of this feature in an application only requires asynchronous communication, which is represented as multithreaded designs in Sockets or as nonblocking primitives in MPI -- MPI_Isend() and MPI_Irecv().
Send/Receive and RDMA
Most of the above capabilities are available through the widely-used two-sided semantics of Send/Receive communication. That is, the communication runs entirely at the user level, allows the local (sending) node to act without copying, and frees the CPU to perform other tasks. Send/Receive does have a drawback though: the remote (receiving) node must copy the message to its final destination; the benefit of zero-copy only exists for the sending node.
With one-sided semantics in remote direct memory access (RDMA), the receiving NIC uses DMA to place the data into a buffer that has been specified by the sending node. RDMA extends zero-copy benefits to the remote node.
For the NIC to access the data through RDMA, the user's page must actually be in memory and not on the disk. Pinning the page to physical memory requires a memory registration, which invokes the operating system. This is actually an expensive procedure as it requires the kernel and is exactly what high-performance networks are supposed to avoid!
Furthermore, the sending node must know the destination memory address on the receiving node. Most applications, such as those written in Sockets or MPI, will require that this information be exchanged prior to communication. The synchronization here is performed through the Send/Receive semantics in a rendezvous protocol, which adds even more overhead.
Workarounds for RDMA
It is possible to overcome RDMA's shortcomings and still realize the benefit of zero-copy communication on the remote node. Certain supercomputers such as the Blue Gene rely on a custom lightweight kernel that only runs one process; because there is no paging, there is no requirement for memory registration.
Alternatively, QsNet works by patching the kernel so that the NIC may access the appropriate data once the page has been loaded into memory. Patches are developed for very specific versions of the kernel based on assumptions regarding the Linux API. Given this level of required specificity, administering a cluster that involves kernel patches can be quite tedious.
As for InfiniBand, it is possible to rely on caching techniques. That is, if a certain memory region will be remotely accessed multiple times, then the software -- an implementation of MPI, for example -- may build a table of memory registrations on the receiving node.
In any case, synchronization remains unavoidable in most programs. The sending node must know the destination memory address on the remote node to perform RDMA. There are some special cases where the address will be known ahead of time, as in MPI-2's remote memory access functions -- MPI_Put() and MPI_Get(). But these routines are not widely used and represent a niche application.
Specific Issues with the Virtual Interface Architecture (VIA)
MVAPICH is a port of MPICH to InfiniBand maintained by D. K. Panda's team at Ohio State University. This implementation provides a reference for other communication layers on VIA-based networks, such as InfiniBand and iWARP. Of particular interest is that OSU's collection of related research papers contain a series of design patterns for software on RDMA networks.
Design patterns are best-practice architecture that permit reuse of a solution to a common programming problem. Some language researchers, such as Paul Graham and Peter Norvig, believe that design patterns are really a sign that the underlying language is incomplete. After all, a pattern implies automation, and automation implies a machine.
By extension, the design patterns from OSU demonstrate that InfiniBand lacks the foundations that would best serve most of its users. Now some designers, such as John Hennessy and David Patterson, believe that an architecture should provide primitives and not solutions. But given that the (committee-defined) InfiniBand standard is over a thousand pages long, it should be fairly obvious which view the IB Trade Association holds.
In contrast, both the Elan and MX libraries (for QsNet and Myrinet, respectively) have been specifically built to present the common functionality required in most applications. The solutions-oriented VIA community should have done the same with their libraries, such DAPL and the OpenFabrics verbs API.
I was motivated to write this article after reading "A Tutorial of the RDMA Model" from IBM's Renato Recio, which in turn was a response to "A Critique of RDMA" from Myricom's Patrick Geoffray. I got the impression that Recio was writing to protect the image of VIA rather than provide a sound rebuttal to Geoffray's technical arguments about RDMA. For example, Geoffray's criticism that RDMA is not adequate for Sockets is met with the response that the user can rely on Extended Sockets or the Sockets Direct Protocol (SDP). Extended Sockets is a different library from Sockets, albeit somewhat similar; SDP is a protocol used above and beyond the RDMA paradigm. Geoffray essentially said that RDMA is handicapped and Recio responded that RDMA has a choice of crutches.
What is particularly telling is that Recio fell back on the old technique of using sales volume to justify technical soundness. He states, "it is interesting to note that almost twice as many new machines in the top100 are using InfiniBand than Myrinet." This is like saying that Titanic was the best movie ever produced since it sold the most tickets. If IBM really did believe the sales-volume pitch, it would stop making POWER chips and simply bundle x86 with its servers.
I wrote this article as a knowledgeable end user; I will leave the marketing brochures to the vendors. At Oxford we used to believe that RDMA was a godsend for the BSP-style programming found in MPI-2 or Cray's SHMEM. Indeed, Geoffray's article states that RDMA networks "can be leveraged successfully for one-sided programming paradigms." After having studied both the paradigms and the networks, I have come to the conclusion that models such as the partitioned global address space languages are really best suited for ccNUMA machines. And indeed, that is what RDMA is: a crude approximation of a non-commodity machine useful only for niche applications.
Sockets work just fine on vanilla Ethernet. MPI works on Ethernet. Google's MapReduce works on Ethernet. Maybe this is the architecture we should be building on.
The author would like to thank Richard Brent and Peter Strazdins for their comments on an earlier draft of this article.
Christopher C. Aycock is wrapping up his PhD from Oxford University, where his thesis topic is in communications programming paradigms for high-performance networks. He is currently a visiting fellow at the Australian National University and can be reached via email@example.com.
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?