Cloud provider Microsoft Azure’s push into HPC continues to gain momentum. In a blog last week, Evan Burness, principal program manager, Azure HPC, announced HC-series Virtual Machine are now available in West US 2 and East US regions. He also highlighted several benchmarking exercises including one in which VMs were scaled to 512 VMs and 22,528 cores.
HC-series VMs feature Intel Xeon Platinum 8168 processors that offer the fastest AVX, AVX2, and AVX-512 clock frequencies from the Intel Xeon Scalable first generation family. Microsoft reports, this enables customers to realize a greater performance uplift when utilizing AVX-optimized applications.
“HC-series VMs expose 44 non-hyperthreaded CPU cores and 352 GB of RAM, with a baseclock of 2.7 GHz, an all-cores Turbo speed of 3.4 GHz, and a single-core Turbo speed of 3.7 GHz. HC VMs also feature a 700 GB local NVMe SSD, and support up to four Managed Disks including the new Azure P60/P70/P80 Premium Disks,” wrote Burness.
“A flagship feature of HC-series VMs is 100 Gb/s InfiniBand from Mellanox. HC-series VMs expose the Mellanox ConnectX-5 dedicated back-end NIC via SR-IOV, meaning customers can use the same OFED driver stack that they’re accustomed to in a bare metal context. HC-series VMs deliver MPI latencies as low as 1.7 microseconds, with consistency, bandwidth, and message rates in line with bare-metal InfiniBand deployments. For context, this is 8x to 16x lower network latency than found elsewhere on the public cloud.”
To see how far HC-series VMs could scale, Microsoft benchmarked it using CP2K. “We chose CP2K for several reasons. For one, it’s widely-used both in academia and industry. In fact, CP2K is one of 13 applications used by PRACE as part of the Unified European Applications Benchmark Suite to drive acceptance testing of supercomputers deployed in Europe,” wrote Burness, also noting that CP2K benefits from AVX-512 and is a good demonstration of what is possible when the latest hardware and software capabilities come together.
One of the benchmarks exercises used the LiHFX benchmark, which is a single-point energy calculation simulating a 216-atom Lithium Hydride crystal with 432 electrons, HC-series VMs successfully scaled to 512 VMs and 22,528 cores. Once again, the return on compute investment is significant. At top-end scale and compared to our baseline of 24 VMs, HC VMs provided a 26.2x improvement in cases-per-day throughput for a 21.3x increase in VM resources.
Intel’s Trish Damkroger, VP & GM of the Extreme Computing, is quoted in the blog, “The Azure HC-series, powered by Intel architecture, gives our customers the high-demand workloads and application performance needed to deliver quicker insights at scale. Our collaboration with Microsoft brings HPC in the Cloud leadership to the forefront, showing performance efficiency and capacity needed for data-intense compute jobs across industries.”
Link to Microsoft blog: https://azure.microsoft.com/en-us/blog/azure-hc-series-virtual-machines-crosses-20000-cores-for-hpc-workloads/