ScaleMP has added two new virtual SMP products to its lineup: a dynamic cloud provisioning offering and a new switchless entry-level product for four-node clusters. Both are based on the company’s vSMP Foundation technology, a software solution that turns x86 clusters into virtual SMP machines.
The new cloud solution is perhaps the more interesting development since it allows x86 servers to be aggregated and dis-aggregated on the fly. Up until now, the ScaleMP technology could only be applied statically. The existing solutions were aimed at customers who wanted to either replace proprietary non-x86 SMP machines, scale up beyond four-socket x86 servers, or turn small clusters into more easily managed shared memory systems. In all cases though, the solution involved plugging vSMP firmware into servers to create a more a less permanent configuration.
By making the vSMP software network bootable, ScaleMP enables cluster workload management software systems to provision their virtual SMP software on top of selected cluster nodes in a large-scale setup, even a private cloud. The software should work with all major provisioning and workload management solutions. The idea here is to be able to change the nature of the cluster architecture dynamically, on a per job basis, enabling the construction of SMP machines when needed, and then the subsequent dismantling of those same machines when applications just require vanilla x86 nodes.
So, for example, chip design companies normally use large server farms to run most electronic design automation (EDA) tools. Most of these are throughput apps, needing lots of individual threads, but not a great deal of memory per thread. Just before tape-out, a routing step is required to figure out the wiring layout. This application tends to be more data-intensive, and while requiring a relatively small number of cores, it needs a large amount of shared memory. Many EDA datacenters end up using “end-of-the-aisle” computers for routing and these tend to be expensive, big memory servers dedicated for this type of workload.
With ScaleMP’s cloud offering, those same EDA datacenters could borrow some of the server farm nodes to temporarily construct virtual shared memory machines to run the routing app. The idea, of course, is to save all the costs associated with installing and running a separate system for this one application, and perhaps make better use of the infrastructure already in place.
Since the cloud version uses the same vSMP technology as the static solution, the maximum SMP configuration is still 16 nodes. Assuming a dual-socket Nehalem EP node, that translates into a core count of 128 and a memory capacity of 4 TB. As Intel’s Nehalem EX processors come into production in a few months, we can expect larger core count and memory capacities to be supported.
The vSMP Foundation for Cloud product is currently available in beta, and will be generally available Dec. 1, 2009. The base price is $65,000, which includes the management node. Each additional node costs an extra $650.
By the way, ScaleMP is not the only company that thinks building shared memory systems on the fly is a good idea. Newcomer 3Leaf Systems also unveiled a similar solution this week, called the Dynamic Data Center Server. In 3Leaf’s case, a custom ASIC plus software is used to make the virtual SMP machines. For an in-depth look at that solution, read our earlier 3Leaf coverage.
The second new offering from ScaleMP is for low-end clusters. The company is unveiling its second generation Direct Connect technology that allows four-node (8-socket) clusters to be aggregated without the benefit of an InfiniBand switch. The original vSMP technology allowed only two nodes to be hooked together without a switch. With Direct Connect 2, users can now hook up four nodes using just cables, with the network routing done as part of the vSMP software. The goal here is to lower the cost of these entry level clusters by as much as 20 percent.
According to ScaleMP founder and president Shai Fultheim, performance is actually improved without the switch in the loop. “We can provide customers anywhere between 30 and 500 percent better performance compared to eight-socket AMD machines that are being offered today by Sun and HP,” he says.
In fact, Fultheim claims a ScaleMP virtual eight-socket machine — a four-node Nehalem EP cluster aggregated by vSMP — will outperform an eight-socket Nehalem EX server. He says although it may seem like an integrated eight-socket node would easily outrun four two-socket nodes connected by InfiniBand, vSMP is able to take advantage of its intelligent memory caching and pre-fetching technology, therefore eliminating a lot of socket-to-socket communication altogether. In addition, Nehalem EP actually offers better memory bandwidth on a per core basis, as well as higher clock frequencies for both memory and CPUs, compared to Nehalem EX.
Given the company’s recent partnership with Cray, we can expect to see the new switchless solution show up in Cray’s CX1 personal HPC machine, especially the four-blade LC configuration. Other partnerships may be in the offing, but since the installation of ScaleMP firmware is relatively straightforward, customers can choose to build these vSMP-equipped mini-clusters themselves.
ScaleMP’s focus on cloud environments and low-end clusters seems to reflect where the company is anticipating the steepest growth. Users with small and moderate sized clusters already represent the majority of ScaleMP’s customer base. Rather than looking to replace high-end SMP machines, these users just want a more easily managed platform, and since vSMP abstracts away cluster management, system operation is simplified. The cloud offering also reflects the same need for infrastructure simplification, but at the other end of the scale. If the technology can keep its promise to deliver virtualization without compromising performance, its place in the HPC ecosystem looks bright.