"GPU Starter Kit" 4U server -- 6000 W, 3ft depth, $100,000 |
Accelerating HP DL380 Servers
HP Proliant series DL380 servers are economical, workhorse machines. In a 2U configuration with dual Sandy Bridge CPUs (16 cores), these servers provide a cost-effective mix of performance, efficiency, and relatively small size. But how to accelerate these machines? There are a number of problems with using GPU or Xeon Phi accelerator boards, including:
- Riser board power consumption limited to 150 W
- Mechanical -- the accelerator, because of its 2 slot width, would consume 2 out of the 3 available riser slots
- Scalability -- GPU and Xeon Phi accelerators are not scalable across platforms, in two ways (i) inability to start with board sizes that fit a given machine configuration, and add as many as possible for that configuration (i.e. poor accelerator granularity), and (ii) inability to support the same applications and code bases across platforms
- Virtualization -- network I/O has to traverse the PCIe backplane, OS, VM layer, motherboard Ethernet -- and then back again -- because network I/O is not located directly on the accelerator. Intel calls this "TCP/IP over PCIe" encapsulation, an euphemistic term for "slow"
HP Proliant DL380p server with CIM accelerator installed. Up to six (6) accelerators can be installed in one DL380p without exceeding 300W total riser limit |
HP Proliant DL380p server with CIM accelerator installed on middle riser. Up to three (3) accelerators per riser can be installed without exceeding 150W riser limit |
Side view of HP Proliant DL380p server with CIM accelerator installed on middle riser. Note the middle slot is used and the CIM card is only one slot width, allow up two (2) more accelerators to be installed |
Side view of HP Proliant DL380p server with Sig A103 card installed on the bottom riser. Note the bottom slot is used and the CIM card is only one slot width, allow up two (2) more accelerators to be installed |