Hi, Pasi,
Here are some throughput numbers we see with disktest, one tcp connection (iscsi session).
The setup are between a pair of chelsio 10G adapters. The target is Chelsio's ramdisk target with data discarded (similiar to IET's NULLIO mode). The Chelsio target is used because of the digest offload and payload ddp.
The ethernet frame is standard 1500 bytes.
The numbers are about 3 months old, but you get the idea :) For cxgb3i driver, since the digest is offloaded the performance is very similar in the digest off case, so only the numbers for digest on are shown.
We will re-run the tests and get the cpu stats too, will keep you posted.
Test cxgb3i iscsi-tcp iscsi-tcp
digest on digest on digest off
(MB/sec) (MB/sec) (MB/sec)
=======================================================
512-read 36.85 34.13 36.69
1k-read 71.91 58.52 66.81
2k-read 137.24 97.75 128.46
4k-read 280.61 137.98 214.04
8k-read 531.34 201.87 325.09
16k-read 953.67 226.49 429.32
64k-read 1099.57 248.57 626.30
128k-read 1102.65 256.04 613.94
256k-read 1105.28 262.28 642.73
512-write 39.54 34.18 38.36
1k-write 79.52 56.51 75.06
2k-write 158.03 84.12 140.85
4k-write 314.56 126.33 282.72
8k-write 559.83 155.49 528.24
16k-write 968.84 168.50 676.38
64k-write 1099.31 182.82 978.82
128k-write 1074.62 182.55 974.18
256k-write 1063.85 185.67 972.88
-----Original Message-----
From: Pasi Kärkkäinen [mailto:pasik@xxxxxx]
Sent: Monday, December 15, 2008 6:06 AM
To: open-iscsi@xxxxxxxxxxxxxxxx
Cc: linux-scsi@xxxxxxxxxxxxxxx; michaelc@xxxxxxxxxxx; James.Bottomley@xxxxxxxxxxxxxxxxxxxxx; Karen Xie
Subject: Re: [PATCH v2 0/2 2.6.29] cxgb3i -- open-iscsi initiator acceleration
On Tue, Dec 09, 2008 at 02:15:22PM -0800, Karen Xie wrote:
>
> [PATCH v2 0/2 2.6.29] cxgb3i -- open-iscsi initiator acceleration
>
> From: Karen Xie <kxie@xxxxxxxxxxx>
>
> Here is the updated patchset for adding cxgb3i iscsi initiator.
>
> The updated version incorporates the comments from Mike and Boaz:
> - remove the cxgb3 sysfs entry for the private iscsi ip address, it can be
> accessed from iscsi.
> - in cxgb3i.txt, added error message logged for not setting MaxRecvDataSegmentLength properly.
> - renamed cxgb3i Makefile to Kbuild
> - removed "select ISCSI_TCP" in Kconfig
> - consistent handling of AHS: on tx, reserve rooms for AHS; on rx, assume we could receive AHS.
> - add support of bi-directional commands for ddp setup,
>
> The cxgb3i driver, especially the part handles the offloaded iscsi tcp connection mangement, has gone through the netdev review (http://marc.info/?l=linux-netdev&m=121944339211552, http://marc.info/?l=linux-netdev&m=121989660016124).
>
> The cxgb3i driver provides iscsi acceleration (PDU offload and payload data direct placement) to the open-iscsi initiator. It accesses the hardware through the cxgb3 module.
>
Hello!
Do you guys have performance comparison/numbers for normal open-iscsi over tcp vs. cxgb3i accelerated?
Would be nice to see throughput/iops/cpu-usage statistics..
-- Pasi
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
