13-Aug-2003

OpenVMS performance improvements

Greg Jordan from HP's OpenVMS engineering team talked about performance improvement work being done in OpenVMS, particularly for the upcoming version 7.3-2.

Some of the highlights include:

  • LAN fastpath, which will move the interrupt load for network devices from the primary processor to a secondary CPU. This will have significant benefits for people experiencing interrupt saturation on the primary CPU. This work also includes fastpath for PEDRIVER, so that NI clusters will benefit.
  • XFC work to increase performance by reducing the amount of code synchronized by spinlock aquisition. This should particularly help with the performance of small block I/O.
  • COPY and SEARCH have both been modified to use 127K buffers instead of 64K buffers. Improvements will be most noticable for copying and searching large files.

Further enhancements are being looked at for COPY and SEARCH so that both utilities use multiple parallel I/O.

Greg also discussed the performance improvements available with the GS1280 machines, and showed real life performance information to back it up.

Although the GS1280s are still NUMA machines, the memory latency is far better than the Wildfire class machines. The Wildfires have an average memory latency for local memory of about 300 nanoseconds, and 900 nanoseconds for remote memory.

The Marvel series is 70 nanoseconds for local memory, and at worst, only 240 nanoseconds for remote memory. This, in combination with the spinlock work that has already been done in OpenVMS 7.3-1, translates into Marvel's scaling far better than any previous machine to run the operating system.

With Marvel, people are having no problems running 16 CPU, where previously the "knee in the curve" was at about 8 to 10 processors.

While OpenVMS Engineering have announced 32 processor support, the scaling is not quite there yet. But they continue to work on it.

Posted at August 13, 2003 10:58 AM
Tag Set:

Comments are closed