NFS sequential read

In Figure 1 a sequential-read FIO workload is running against a file system using NFS-connected shares. The workload is scaled from 1 to 128 jobs running on the KVM guest. The storage space is provided from the protocol server contributing an additional 64 GiB of page cache.

Similar to the observation above, high levels of throughput are observed with a workload of 1 and 2 jobs, as the KVM guest page cache can contain the total file size being used for the measurement run. The CPU load level on the guest and host closely tracks the number of workload jobs at 1 and 2 jobs. The close tracking of CPU load appears related to satisfying the I/O request from the guest page cache.

Starting at 4 jobs, the total file size exceeds what that page cache can hold, leading to a substantial drop in throughput as page reclaim and read requests from the virtual device increase.