RSC Packs a Petaflops of Xeon Phi in a Rack
RSC Group, a four-year-old maker of clusters based in Moscow that has created some of the most energy efficient machines installed in Russia, got plenty of attention last week at SC13 with its new PetaStream system.
Like many top-end systems created today, the PetaStream is a hybrid design, and in this case it mixes a modest number of Xeon processors from Intel with a much larger number of Xeon Phi X86 coprocessors to get more than 1 petaflops of aggregate double-precision number-crunching power into a single rack.
The PetaStream system cheats a little bit on the server density by using a non-standard rack that is a meter wide (3.28 feet). A standard server rack is 19 inches on the inside and usually around 24 inches (two feet) on the outside. That said, the PetaStream system still points the way to the kind of dense packaging of compute capacity that will be necessary to approach exascale compute capacities in the next decade.
The PetaStream system has an Intel Xeon E5-2690 v2 processor as the basis of each node in the machine. This is the new ten-core “Ivy Bridge-EP” processor, which was announced by Intel in September and which clocks at 3 GHz. RSC is using Intel’s S1600JP motherboard, which was designed for single-socket, half-wide server nodes but which can accept the E5-2690 v2 chip, which was made for two-socket machines. Each one of the Xeon E5 processors has eight Xeon Phi coprocessors hanging off its PCI-Express bus. The system has an embedded version of the Xeon Phi that has 60 cores and 240 threads on each “Knights Corner” chip, which delivers over 1 teraflops of peak performance on double-precision calculations. This Xeon Phi card has 8 GB of GDDR5 memory and runs a skinny Linux operating system for workloads that are offloaded from the CPU.
Customers can use Intel True Scale or Mellanox Connect-IB adapter cards on the server nodes to link out to InfiniBand or Ethernet interconnects. With the two PCI-Express 3.0 x16 slots, RSC suggests that four InfiniBand ports running at 40 Gb/sec (QDR) or 56 Gb/sec (FDR) coming out of the module is a reasonable configuration where high bandwidth and low latency are important.
The Xeon and Xeon Phi components are packaged up in a compute module that has direct liquid cooling on all of the hot components in the module. Either water or glycol can be used as the coolant for the machine, according to Alexander Moskovsky, CEO at RSC Group. The liquid cooling system is designed to allow as much as 400 kilowatts of electricity to be consumed by the rack’s components and then pulled off as heat.
The PetaStream modules have room for five solid state drives for local storage on the cluster, and RSC has chosen Intel’s DC S3700 or DC S3500 series drives for the PetaStream machine, which have 800 GB of capacity. With all 640 SSDs in the rack, that gives 512 TB of storage.
The PetaStream system rack slides compute modules in both the back and the front of the rack, for a total of 128 modules. That yields 1,024 Xeon Phi coprocessors for a total of 61,440 cores and 245,760 threads, all crammed into a rack that is 7.2 feet tall and measures 3.28 feet on a side.
“Our calculations show that one rack of RSC PetaStream is close to eight racks of X86 servers,” Alexey Shmelev, chief operations officer and co-founder of RSC, explained at the launch of the machine at SC13 last week. “This is not the measurement of peak theoretical performance, but the measurement on real applications.” And, he added, the PetaStream system required half of the energy as the X86 machine to solve the same problem.
Because the machine can be configured in a number of different ways, RSC was reluctant to provide a base price for the PetaStream rack. But Moskovsky said that the Xeon Phi cards were a big part of the system cost, which stands to reason given that there are 1,024 of them in the rack. When pressed about how the PetaStream machine would compare to a plain vanilla cluster of X86 servers, Moskovsky said that that the “flops per dollar would be price competitive” and left it at that.
The idea that is espoused again and again in cluster designs with compute offload, liquid cooling, or both is that the benefits of these technologies pay for themselves in savings in space, power, and cooling. That is true so long as a customer has the kind of software that can be tweaked to offload routines to the Xeon Phi coprocessor – and the time to make such changes to their code.
The PetaStream machine will be even more interesting when Intel ships its next generation “Knights Landing” Xeon Phi chip, which Intel talked a bit more about at SC13 last week. This Xeon Phi chip will have its own local memory on the die as well as DDR4 memory on the package; it will also plug into its own socket on a system board and will be available as a standalone compute module that no longer needs a Xeon processor telling it what to do over the PCI bus.
The roadmaps that we have seen show the Knights Landing chip delivering around 3 teraflops of double-precision math and coming to market in 2015 or so. In theory, a PetaStream rack could triple up to 3 petaflops in its rack without too much effort. It looks like the Knights Landing chip will have about the same thermal envelope as the current Knights Corner Xeon Phi chips, so putting a larger number of Xeon Phi chips into a PetaStream rack is probably not going to be easy even if it is possible. Then again, that is what engineering is all about, and RSC has time to figure out how to cram even more compute into a rack.