Hi Everyone,

A while back I posted a thread on iSCSI and Jumbo frames.

I figured this was worth a new thread as I have plenty of new info based
on the testing I have done.

Perhaps others can provide their findings and help me figure out why jumbo
frames is slower for me ! :-)

All testing done with TSATEST running 40 plus iterations !

Here is my test setup:

Server A - iSCSI Initiator
Acer R710 (Intel dual Xeon board) with 1 3.0GHz/800FSB CPU
Hyperthreading off
2GB Dual Channel DDR2
onboard Intel Gigabit
NW65 install from SP3 overlay
NW65OS3b server.exe update
N65NSS3a nss update

Switch - Cisco 3750G switch with jumbo (9018 byte frame) turned on.

Server B - iSCSI target
HP Dl380 G3 with dual Xeon 3.0 / 533 FSB
Onboard Broadcomm Gigabit
NW65 install from SP3 overlay
NW65OS3b server.exe update
N65NSS3a nss update

STORAGE - IBM EXP300 (Ultra2) enclosure with 6 of 36GB 10K SCSI drives
connected to the HP using the onboard RAID in the DL380. RAID 5 config.

DATA - 3 zenworks workstation image files totally 14.7GB in size.

Test 1 - Server B mounts the volume directly and run TSATEST
Result 1 - 2918MB/Min (48.6 MB/sec)
Comment 1 - gives us a measure of the actual disk performance.

Test 2 - Setup ISCSI from A to B, recreate volumes and mount on Server A.
Result 2 - 1208 MB/ Min
Comment 2 - This is all default settings including 1500 byte MTU.

Test 3 - As per 2 but with "set TCP minshall algorithm = on" for both
servers. (nagle and delayed ack are on by default)
Result 2 - 1320 MB / Min
Comment - small but consistent increase in performance.

Test 4 - As per 3 but with Nssstart.cfg on "A" from TID 10093351 (except
for the readaheadblks setting).
Result 4 - 1174 MB / Min
Comment 4 - Seems this TID doesn't suit this scenario. The defaults are
better !

Test 5 - As per 3 (reversed the nssstart.cfg) but with "jumbo=9000" on the
lan card, "Maximum Physical Recieve Packet Size = 16400" and "MTU
Blackhole Detection and Recovery = on" on both Server A and Server B.
Result 5 - ~895MB/Min
Comment - What the ????

Test 6 - Updated CE1000B on "A" to current (7.84) and Q57 on "B" to
current from HP site (8.56 - had to "force install"). Dropped "jumbo"
to "jumbo=4114" at both ends.
Result 6 - 866 MB / Min.
Comments - confirmed that Minshall was still on. Scratched head....

I intend to update TCPIP and Wsock next... stay tuned !