Ron Gordon
Director – Power Systems
Enhancing the performance of systems means better customer support, lower costs, and the ability to invest in solutions needing very high performance (i.e, analysis of streaming data). Historically, the enhanced performance was achieved by getting faster speed infrastructure, with higher speed cores or adapters, or by investing in a clustered environment of multiple servers and storage. While clustering is still being implemented frequently today, the “Moore’s Law” says processor speed is possibly reaching a maximum. Seeing the cost of clusters and maximum processor speeds, IBM Power Systems is creating new capabilities and architectures to support “accelerators” (i.e., co-processors or additional special purpose processors or ASICs).
Following is a quick review of several of these:
NVMe flash memory/storage: Non-Volatile Memory Express is the next speed bump in SSDs and can be viewed as extended memory or a storage device accessed as a LUN. The performance of an NVMe adapter, or disk, is 7x read IOPS and 2x write IOPs over SSDs, and 50% more read IOPs and 20% more write IOPs than Flash. The NVMe adapters for Power can be up to 3.2 TBs and are supported by Linux applications. Applications that can benefit from this are read intensive applications, caching, and very fast restart solutions.
GPU: The name, Graphic Processing Unit, may be misleading. A GPU is really a vector processor, or parallel processing unit, that can be used in many applications, such as geospatial, video displays, map reduction, data analytics, etc. Normally, a GPU is attached via a PCIe link and thru programming data. Instructions are then downloaded to the GPU, which contains thousands of processors and memory, and processing is done in parallel, providing extremely fast results to iterative processes. This also has the side benefit of freeing up the main system processor, for additional threads to be executed in the main program. In the latest Power 822LC model, an advancement has been made where 2 or 4 GPUs are on the system board and connected to the processor, via a new technology called NVlink. This allows data to be transferred to and from the GPU at 80 GB/s, as opposed to 16 GB/s on the PCIe bus. Net result is high performance.
FPGA/ASIC: Field Programmable Gate Arrays and Application Specific Integrated Circuits are adapters with processor(s) and memory. Through programming, they can perform specific tasks, such as device driver execution, access to storage, computations, etc. As such, these offload the main system processor and can perform repeatable tasks very fast. In Power Systems, PCIe attached FPGAs are supported, and Power Systems has created a technology called CAPI (Coherent Application Programming Interface) which speeds up the FPGA application, by providing a direct coherency to the processor. Today, Power Systems has enabled applications to use the CAPI/FPGA adapter for use in in-memory databases, data compression, statistical processes, and video processing. CAPI and the Power Systems adapters are Linux supported.
Let Mainline assist you in a deeper analysis of how accelerators can assist your business needs thru Power Systems and new technology.
Please contact your Mainline Account Executive directly, or click here to contact us with any questions.