Re: port-amd64/57266: Poor disk performance on VirtualBox vioscsi0 and NetBSD 10.0_BETA

[mlelstv%serpens.de@localhost (Michael van Elst)]

The following reply was made to PR port-amd64/57266; it has been noted by GNATS.

From: mlelstv%serpens.de@localhost (Michael van Elst)
To: gnats-bugs%netbsd.org@localhost
Cc: 
Subject: Re: port-amd64/57266: Poor disk performance on VirtualBox vioscsi0 and NetBSD 10.0_BETA
Date: Sun, 12 Mar 2023 11:02:08 -0000 (UTC)

 bbartlomiej.mail%gmail.com@localhost (Bartek Krawczyk) writes:
 
 >It shows:
 >messages:Mar 12 11:06:11 netbsd /netbsd: [   1.0546101] vioscsi0: 
 >cmd_per_lun 128 qsize 1024 seg_max 126 max_target 2 max_lun 1
 
 If the possible 128 commands get queued, each should take at most 8
 queue entries (but with 64k blocks and 4k pages it can take 18 (or 17?)).
 
 >logs didn't show any errors but unpacking pkgsrc also took ~30min. The 
 >VM has 12 vCPUs (Ryzen 9 3900x) and ~16GB of RAM assigned.
 
 Unpacking pkgsrc can vary in time, I get natively:
 
 - 42s to a journaled ffs on NVME.
 - 10m to a not-journaled ffs on SATA HDD.
 
 So 30m is slower, but on a virtualized disk maybe not that much.
 A slow disk also makes the queue larger (and the risk to overflow it)
 which makes it even slower.
 
 I tried on qemu with vioscsi, but the queue was never large and unpacking
 pgksrc never overflowed it.
 
 [   1.0000070] vioscsi0: cmd_per_lun 128 qsize 256 seg_max 254 max_target 255 max_lun 16383
 
 
 To get some information on how fast your emulated disk is, you could
 compare the unpacking (lots of transactions) with reading the raw
 disk sequentially (bandwidth matters) using e.g.:
 
 dd if=/dev/rdk0 of=/dev/null bs=1024k count=10240