[OpenIndiana-discuss] ISCSI SSD Performance
Grant Albitz
GAlbitz at Albitz.Biz
Sun Feb 17 00:17:02 UTC 2013
Setting the MTU to 1500 did not affect read speeds but it cut my write speeds in half. Flow control on or off did not make any difference.
The cable is somewhat ruled out because I am seeing the same exact performance on 2 different esxi hosts.
-----Original Message-----
From: Jim Klimov [mailto:jimklimov at cos.ru]
Sent: Saturday, February 16, 2013 7:04 PM
To: openindiana-discuss at openindiana.org
Subject: Re: [OpenIndiana-discuss] ISCSI SSD Performance
On 2013-02-17 00:52, Grant Albitz wrote:
> Yes jim I actually used something similar to enable the 9000 mtu that's why I want familiar with the config file method.
>
> dladm set-linkprop -p mtu=9000 InterfaceName
>
>
> Flowcontrol is currently off on the zfs host but enabled by default on esxi, I am going to try enabling flow control first and if that doesn't improve anything I will disable it on both ends.
I'd also google for a way to verify that MTU=9000 frames are actually used by both sides and that packets are not fragmented into smaller frames and reconstructed with more overhead that would be needed if they originally were smaller.
I don't think I've tested if dladm can enforce larger mtu's when the NIC driver config file doesn't have it; I do know that "ifconfig mtu"
settings did fail unless I fixed up the e1000.conf (in my case) and rebooted.
Also, since you use point-to-point cable connections, you might use largest MTUs - like 16k, if supported by NICs on both sides.
Just in case, to also validate (perhaps replace) the cable which connects the machines. It may be possible that some strands have worse connectivity than others, causing jitter and retries in only one direction. "netstat -s" might help to diagnose that...
Good luck,
//Jim
_______________________________________________
OpenIndiana-discuss mailing list
OpenIndiana-discuss at openindiana.org
http://openindiana.org/mailman/listinfo/openindiana-discuss
More information about the OpenIndiana-discuss
mailing list