Re: XEN 4.11 and 9.99.48 DOMU performance

[Frank Kardel <kardel%netbsd.org@localhost>]

Hi Greg!

That was my first guess, but DOM0 is, as I wrote mostly idle in this 
state. If it where the bottleneck we should

expect it to be busy doing IO, or not?

On 03/10/20 16:50, Greg Troxel wrote:
> I wonder if the issue is that the dom0 is single threaded, serializing
> all the IO.
>
> Have you done dd bs=1m from the dom0 raw disk, the dom0 backing file,
> and the domU ('raw disk')?   I used to see X, 0.9X, and (0.9)^2 X, ish,
> on somewhat old Xen and older NetBSD.
Before this load test I reach 990Mb/s in DOMU instead of 1100Mb/s on DOM0.

Now is this situation it joins the waiting game:
system# dd if=/dev/zero of=file bs=1m count=1024
[ 351633.9327690] load: 1.09  cmd: dd 20949 [biowait] 0.00u 0.00s 0% 1728k
[ no stats from dd yet ... ]

> I would run "systat vmstat" in the domU, to see the disk busy fraction
> and transfer rate.
Mostly 0. spurious spurts of io up to 90MB/s, but most if the time nothing.
> I also wonder if there is some sort of fsync in the domU leading to
> virtual FUA leading to fsync in the dom0 going on.
Could be, but in that case I would still expect a reasonably data stream 
to the disks then and not
several seconds of inactivity.

Frank