next up previous
Next: Other Factors Up: The Kernel-Space NFS Server Previous: Moderate Load

Heavy Load

For the heavy load, we set the initial load to 800 and increment the load by 800 at each run:

Run # of Server # of # of Processes Load Throughput Response Time
  Threads Clients Per Client (Ops/Sec) (Ops/Sec) (Msec/Op)
1 1 5 4 800 356 56.6
2 1 5 4 1600 212 97.8
3 1 5 4 2400 162 127.5
4 1 5 4 3200 138 151.9
1 2 5 4 800 569 36.8
2 2 5 4 1600 381 54.8
3 2 5 4 2400 311 67.5
4 2 5 4 3200 264 78.6
1 4 5 4 800 782 27.4
2 4 5 4 1600 594 35.9
3 4 5 4 2400 494 42.8
4 4 5 4 3200 422 49.8
1 8 5 4 800 799 6.9
2 8 5 4 1600 1125 19.4
3 8 5 4 2400 907 23.9
4 8 5 4 3200 775 27.7
1 16 5 4 800 801 6.1
2 16 5 4 1600 1539 14.2
3 16 5 4 2400 1190 18.9
4 16 5 4 3200 967 22.8

It is quite clear that under the heavy load, the NFS server performance benefits tremendously when the number of the server threads was increased from 1 to 8. When it is above 8, the performance gain was not as significant as the lower thread count.

Overall, the SFS benchmarks show that

For our specific server hardware setup, 8 kernel NFS server threads strike a balance between performance and resource usage.

To further study the Linux NFS server performance, we need to run the benchmarks on


next up previous
Next: Other Factors Up: The Kernel-Space NFS Server Previous: Moderate Load
H.J. Lu
1999-08-15