Solutions for Ultra-Low Latency, High Throughput Computing

Solutions for Ultra-Low Latency, High Throughput Computing

Solutions for Ultra-Low Latency, High Throughput Computing

Specialized computing tasks often demand network and I/O performance beyond those of typical data center applications. Workloads such as high-frequency trading require ultra-low levels of latency and jitter. high performance computing (HPC) requires high throughput and low latency.

Today’s network stacks based on Linux* and Microsoft Windows* are generally not well optimized to meet these requirements. Intel® Ethernet Controllers and network adapters have been well received in the financial services and high performance computing (HPC) communities for their exceptional reliability, performance, and I/O virtualization capabilities. This report shows how financial services and high performance computing (HPC) customers can benefit from the use of Intel® Ethernet Controllers and adapters in low-latency and high performance applications.

MCorelab has reported latency results using its MCoreRT* parallel processing software platform that are very close to the theoretical minimum latency of the Intel® Ethernet Converged Network Adapter X520 (Intel® Ethernet CNA X520). A team at Intel’s Jones Farm Performance Lab undertook testing to verify those claims using the standard methods and processes that Intel uses to quantify the performance of Intel® Ethernet products. MCoreRT* incorporates the following optimizations to support ultra-low latency computing on Intel® architecture-based hardware:

  • MCoreRT’s kernel-bypass network stack allows applications to directly access the network hardware. This capability avoids the overhead of the scheduler and other OS mechanisms, providing for minimum latency and maximum throughput.
  • MCoreRT’s scalable I/O and event-processing system works to optimize the Intel® platform by scalable utilization of multi-core resources, seamlessly feeding I/O and event streams to the processing cores and providing ultra-fast and scalable event processing and inter-processor communication to applications.
  • Processor quiescence analytics proactively analyze all available processor cores within the system to help determine which core or cores are experiencing the least noise and interrupts. MCoreRT* helps to identify and assign threads to specific processor cores on that basis, minimizing latency and maximizing throughput.

Videos >