On Mon, Nov 26, 2018 at 3:22 PM Nathan Stratton Treadway <natha...@ontko.com> wrote: > > On Mon, Nov 26, 2018 at 14:15:41 -0500, Chris Nighswonger wrote: > > That makes more sense. I've set it to 10 and inparallel to 2 and we'll > > see how it goes tonight. > > > > On Mon, Nov 26, 2018 at 2:01 PM Cuttler, Brian R (HEALTH) > > <brian.cutt...@health.ny.gov> wrote: > > > > > > I believe maxdumps is max concurrent dumps across all clients. > > > > > > You might have 10 clients each with an inparallel of 2, giving 20 > > > possible concurrent dumps, but because of server limitations you > > > might set maxdumps to something between 2 and 20. > > Actually, it is the other way around: when you kick off amdump, the > driver retrieves the "inparallel" configuration option and then starts > up that number of dumper subprocesses -- so "inparallel" controls the > total number of dumpers available for the whole run. > > In contrast, the amanda.conf man page section on maxdumps says: "The > maximum number of backups from a single host that Amanda will attempt to > run in parallel." So that controls how many (out of the "inparallel" > total dumpers available) might simultaneously run on one host at the > same time. > > (Maxdumps defaults to 1 to avoid two dumpers on the same client > competing with each other for disk I/O [or worse, thrashing the disk > heads by interweaving reads from different places on the device], CPU > [e.g. for compression], and/or network bandwidth. This seems likely to > be a reasonable default, though I guess if you had multiple physical > disk devices each of which was relatively slow compared to CPU and > network bandwidth then it would make sense to run more than one dump at > a time. But I'd do careful testing to make sure that's not actually > slowing down the overall dump.)
So I've set these params back as they were before (ie. inparallel 10, maxdumps 1) and opened the throttle on the bandwidth. We'll see if things get better or worse tonight. I have a few more DLEs I need to split which may help the over all problem.
