Re: [Scst-devel] iSCSI-SCST Performance

[Alexander T. <ale...@gm...>]

Bart hi,

thanks for the tips.

Reading through the scst readme file on the measure_latency part (and not
being an expert by any means..) am not  really sure on how to enable it.
Its not clear to me. Would I just add the measure_latency in my scst.conf
file to enable it and then use it to measure??

Copying from the readme file:
echo 1 >/sys/kernel/scst_tgt/measure_latency (OK) sleep 10 # Wait until an
initiator has submitted multiple I/O requests (OK) cat
/sys/kernel/scst_tgt/targets/ib_srpt/fe80\:00.... (directory
/sys/kernel/scst_tgt/targets/ib_srpt does not exist)
In my case I have the /sys/kernel/scst_tgt/target/iscsi directory.

I am looking through the statistics in /sys/kernel/scst_tgt/sgv and will
let you know.

many thanks.
Alex


On Tue, Feb 5, 2019 at 6:05 AM Bart Van Assche <bva...@ac...> wrote:

> On 2/4/19 6:08 AM, Alexander Turner wrote:
> > I'm using Debian 9 with iSCSI-SCST to present to a windows server using
> > vdisk_blockio (with ZFS). Windows server has 32gb Ram and Intel i7
> > (i3770k) CPU. The Debian server has 32gb ram and an Intel i5-7600k cpu.
> >
> > Both servers have a Mellanox ConnectX2 (10 Gigabit Ethernet) and they
> > are connected directly, e.g. there is no router in between.
> >
> > I'm in the testing phase and getting very poor results, as in:
> >
> > A simple pool (raid 0) of 4 Samsung 840 EVO Drives 1TB each, where I get
> > less than 100mbytes/sec ??
> >
> > my scst.conf looks like this:
> >
> > HANDLER vdisk_blockio {
> >          DEVICE disk01 {
> >                  filename /dev/DebianZFSPool/xDisk
> >                  nv_cache 1
> >          }
> > }
> >
> > TARGET_DRIVER iscsi {
> >          enabled 1
> >          TARGET iqn.2017-08.DebianZFS.TestServer:tgt {
> >
> >                 LUN 0 disk01
> >                 enabled 1
> >
> >                 allowed_portal 192.168.1.110
> >
> >          }
> > }
>
> The numbers in your e-mail show that there is a bottleneck somewhere.
> Have you already tried to determine where the performance bottleneck is?
> Have you considered to enable latency measurements in SCST (see also the
> measure_latency documentation in scst/README)? That will tell you
> whether or not the bottleneck is in ZFS. You may also want to have a
> look at the statistics in the /sys/kernel/scst_tgt/sgv directory. That
> will give you information about the average queue depth.
>
> Bart.
>