Difference between revisions of "Cluster:llk-siam-pp06-abstract"
| Line 12: | Line 12: | ||
Network latency continues to be a bottleneck for computational science | Network latency continues to be a bottleneck for computational science | ||
| − | applications running on Beowulf clusters | + | applications running on Beowulf clusters. First, we develop technology for |
making nanosecond measurements | making nanosecond measurements | ||
between timing points within the Linux kernel. Using this technology, in | between timing points within the Linux kernel. Using this technology, in | ||
Revision as of 13:53, 30 September 2005
Lowering Network Latency in the 2.6 Linux Kernel (v0.1)
Network latency continues to be a bottleneck for many computational science applications running on Beowulf clusters, that is clusters built out of commodity hardware and open source software. First, using open source building blocks, we develop technology for accurately making nanosecond measurements between timing points within the Linux 2.6 kernel. Using this technology, in conjunction with low-level and application benchmarks, we establish that there is measurable network latency in the kernel. With this benchmarking framework in place we examine and evaluate a range of existing software based approaches to reducing network latency in the kernel e.g. the scheduled transfer protocol (STP) and zero copy packet transfer. Finally, we develop and benchmark a really nifty, non-obvious, and groundbreaking approach to reducing network latency in the 2.6 Linux kernel.
Notes
- Way too long right now (129 words)
- Note that there was no obvious way to get nanosecond kernel timing events without engineering it ourselves. "In order to accurately measure where the time was being spent it was necessary to design and develop, using open source building blocks, a timing harness with nanosecond accuracy for the 2.6 Linux kernel." Or something like that.
Lowering Network Latency in the 2.6 Linux Kernel (v0.2)
Network latency continues to be a bottleneck for computational science applications running on Beowulf clusters. First, we develop technology for making nanosecond measurements between timing points within the Linux kernel. Using this technology, in conjunction with other benchmarks, we establish that there is measurable network latency in the kernel. With this benchmarking framework we examine a range of existing software based approaches to reducing network latency in the kernel e.g. the scheduled transfer protocol (STP) and zero copy packet transfer. Finally, we develop and benchmark a method to reducing network latency in the Linux kernel.
