Comparison sequential read

Edit online

In Figure 1 a sequential-read FIO workload is running against a file system. The workload is scaled from 1 to 128 jobs. Values for each of the protocols (iSCSI, NFS, SMB) are plotted. The throughput plot is normalized to 1 SMB job.

Figure 1. Comparison - sequential read
Comparison - sequential read

1 through 2 workload jobs:

  • As described in prior sub-sections, the behavior here is affected by the KVM guest page cache.
  • SMB has the highest throughput, followed by NFS. iSCSI has the lowest throughput, which appears affected by the I/O stack overhead for iSCSI I/Os on the KVM host vs the NFS/SMB protocols.
  • The CPU load is the same for all three attachment types.

4 workload jobs:

  • The relation of the performance from the various protocols stays the same, but the total throughput takes a severe hit when the page cache is exceeded.
  • Increased network utilization, starting at 4 workload jobs, appears to affect throughput levels.
  • As indicated in the previous section, SMB shows the highest CPU load.

8 through 128 workload jobs:

  • The iSCSI, NFS, and SMB results are closely clustered. iSCSI and SMB exhibit a small throughput advantage over each other in a few workload jobs.
  • NFS shows a slight advantage in throughput efficiency. SMB has the lowest throughput per CPU, trending lower with the increased number of workload jobs.