ART@ACC.ARPA (Art Berggreen) (09/09/85)
First of all, this is a VMS question, so those not interested may skip. While testing a third party's network software on a VAX 8600 we experienced thoughput roughly equivalent to our VAX 750. Since our network interface had significantly higher capacity than the measured throughput, we were somewhat surprised. The 8600 was dedicated for the testing and the CPU was 90% idle during the transfers. The key was that the disks were attached to an HSC50 out on the CI bus. After some discussion with DEC, it was determined that there is a limit of about 100 requests/second on an HSC50. If the HSC50 is being read one block at a time, there is a max disk I/O rate of about 400Kbits/sec. The HSC50 does best with large contiguous block transfers which can approach the speed capacity of the disk drive. It appears that the network software causes RMS to read the disk in small chunks (probably 512 or 1024 bytes) causing disk I/O to be a bottleneck. Assuming that the software may not be changed, I was trying to think of other ways to improve disk I/O performance. Having never studied RMS tuning, I am interested if anyone who might have has any ideas. It seems that given appropriate parameters, RMS should be able to read- ahead (also write-behind) and use large requests in the process. Anyone have any ideas or experience? We didn't note what the SYSGEN or PROCESS RMS parameters were on the test machine, but assume that they were fairly standard defaults. This also manifests itself under DCL COPY. Copying a file to NLA0: took TWO AND ONE HALF TIMES LONGER than copying to a new file. Apparently since NLA0: is a sequential device, copying is one block at a time, while file to file copying uses larger blocks. Since we don't run a CI in-house, we can't easily test our ideas. "Art Berggreen"<Art@ACC.ARPA> ------