Dear NFS fellows, We was running bunch of tests where a high latency network have was simulated. Though overall result did much our expectations, I was surprised by iowait reported by the kernel. When we delay network packets from DS by 50ms, top and other tools show high IO waits. However, when MDS returns multiple times NFS4ERR_LAYOUTTRYLATER, to LAYOUTGET, then IO waits are not reported (see attached screenshot). Of course, this has no impact on a running application, but provides wrong information to monitoring system. If for whatever reason MDS always reports NFS4ERR_LAYOUTTRYLATER, then client will show no CPU utilization and such situations won't be noticed. Best regards, Tigran.
Attachment:
Screenshot from 2019-09-19 11-50-35.png
Description: PNG image