Re: Network very very slow... was iSCSI and jumbo frames

[BERTRAND Joël <joel.bertrand%systella.fr@localhost>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA512

Greg Troxel a écrit :
> 
> BERTRAND Joël <joel.bertrand%systella.fr@localhost> writes:
> 
>> I have tried to mount on NAS a NFS volume exported by NetBSD.
>> From NAS, dd if=/dev/zero of=test count=1000 bs=1M shows a
>> throughput to NetBSD server less than 7 MB/s.
>> 
>>> Can you somehow write to the NAS's /dev/zero over NFS?
>> 
>> I'm not sure it's possible.
> 
> Sorry, I was confused.
> 
> What I am trying to say is to start with the machine with the
> actual disk, try reading/writing from the disk, to see what the
> capacity is, and then try iSCSI, working outward one step at a
> time.  You could try mounting the disk iSCSI from netbsd to the
> same netbsd machine, and to other netbsd machines.

	I have done some other tests.

	I cannot disconnect this NAS to reconnect on a Linux workstation (I
have to reconfigure whole qNAP firmware, in particular network
configuration). Thus, I have disconnect NetBSD iSCSI initiator and I
have reconnect NAS volume on a Linux workstation, but through NetBSD
server that acts as a router.

NAS---(wm0)---NetBSD---(agr0)---switch----Linux workstation.

	To be honnest, there are a lot of other _diskless_ workstation on
this network (NetBSD, Linux and FreeBSD).

	I have mounted iSCSI volume on hilbert (i9/32GB, Linux Debian,
Realtek RTL8111/8168/8411, mtu 1500):

root@hilbert:/mnt/bacula# mount -r -t ufs -o ufstype=ufs2 /dev/sdb1 /mnt

	I cannot write on this volume as Linux only mounts FFSv2 partition in
readonly mode. Now, I try :

root@hilbert:/mnt/bacula# dd if=Full-0218 of=/dev/null bs=10M
5119+1 enregistrements lus
5119+1 enregistrements écrits
53687074523 octets (54 GB, 50 GiB) copiés, 505,046 s, 106 MB/s

CPU utilization is about 30% on NAS and NAS indicates a throughput
greater than 100 MB/s. NetBSD shows 900 mbps on wm0 interface.

	I do the _same_ test from NetBSD server :
legendre# dd if=Full-0218 of=/dev/null bs=10m
...
and I wait, I wait...

iftop returns :
                191Mb           381Mb           572Mb           763Mb

└───────────────┴───────────────┴───────────────┴───────────────┴───────
────────
legendre.systella.fr       => euclide.systella.fr  465Kb   460Kb   482Kb
                           <=                     78.5Mb  78.1Mb  81.4Mb

	Thus, NetBSD initiator only reaches 1/10th of the throughput obtained
by Linux. Bottleneck is iscsi initiator implementation on NetBSD side.

	Best regards,

	JKB

-----BEGIN PGP SIGNATURE-----
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=sjeI
-----END PGP SIGNATURE-----