Re: [Scst-devel] Strange performance behavior on top of encrypted raid

[Göran B. <goe...@gm...>]

Am 08.06.2019 um 23:26 schrieb Bart Van Assche:
> I'm using dm-crypt myself on my development workstation. The
> performance that fio reports for the underlying (old) SSD is 52 K IOPS
> for a block size of 4 KB. With dm-crypt on top of a file on top of
> that SSD I see 13 K IOPS. The command I ran to measure random I/O
> performance is as follows:
>
> fio --group_reporting --gtod_reduce=1 --bs=4096 --ioengine=libaio
> --rw=randread --buffered=0 --numjobs=8 --iodepth=128
> --iodepth_batch=64 --iodepth_batch_complete=64 --thread --norandommap
> --loops=2147483648 --runtime=60 --invalidate=1 --name=/dev/sda
> --filename=/dev/sda

Sorry Bart, I fail to see where this is going to.

I am aware that one can't expect the performance of an encrypted volume
to be equal to a non-encrypted volume.

What I was wondering why the overall performance in writing large files
(which should result in sequential writes) in conjunction with blockio
is so much slower compared to writing medium sized files (which should
also result in sequential writes). Also why is fileio (with it large
write buffers) in both cases even faster.
These are the facts I observed.

The iscsi drive is used for backup purposes. there is no high io-depth
nor many threads.
I can't do anything related to dm-crypt, except to check that the cpu is
capable of encrypt/depcrypt at a reasonable speed.
I could try to tweak the raid and test different stripe sizes, but it
worked well as it is before I used encryption.

So there is only SCST where I can change settings.
If there nothing I can do or you think could be the reason for this
strange behavior. Well then I have to make up my own mind on how to
handle the issue.

I am really thankful for the time u put in here.

Göran