Poor speed so far on new servers...

[trini0]

Version 5.0 is out.
I'm eagerly awaiting to see what numbers you got with the new version.
Thanks

[TomHelp4IT]

Ok, sorry about the delay but decided to wait until I got hold of Starwind 5.0, I have just set it up and ran a preliminary test. Setup is as follows:

Starwind SAN server: HP DL80 G6, Windows 2k8 Std, 2GB RAM, HP P410 RAID w/512MB BBWC, 6x Western Digital RE3 SATA drives RAID1+0, 4x HP Gb NICs

Test initiator server is the same, they will be an HA pair when I'm finished. Servers are linked by a CAT5e crossover cable, each NIC set for 9k jumbo frames and Starwind registry optimisations added.

Created a 50GB image file on the RAID10 volume, default Starwind settings, no caching. Mounted it on test initiator server using MS iSCSI Initiator as a simple volume, default formatting.

Tested using IOmeter, 4GB test file, "max read" - 32Kb transaction size, 100% sequential reads, 10 second ramp up time then 120 seconds testing:

Total IOs/second = 3240
Total MBs/sec = 101
Average response time = 18.5ms

Interestingly watching the network usage during the test file creation it averaged only 25%, but this went up to 85% during testing, as the results indicate it is pretty much maxing out the interface. This RAID array will do stupidly fast sustained reads so has no trouble saturating the NIC.
Will try and do my other disks benchmarks later and post the results, also have a pair of SSDs in this server which should be interesting, although one keeps cutting out atm.

[anton (staff)]

Well... 32KB IOs are far from the best choice if you want to get maximum possible through output. Windows itself breaks all file system requests to 64KB chunks (famous MM_MAXIMUM_DISK_IO_SIZE) at class driver and all modern I/O intensive apps (like f.e. Oracle database or video editing stuff) deal with RAW file system and issue multi-megatyte requests. Even with 64KB-128KB ones you should reach GbE limit pretty fast (~125MB/sec in single direction, read or write).

In any case I'm assuming you have NO performance issues with V5 of StarWind? Even with cache effectively DISABLED?

[TomHelp4IT]

Not sure if I mentioned it before but I use the IOmeter settings suggested in this forum: http://communities.vmware.com/thread/73745 - it seems to be the only decent comparison of SAN performance that I have found, hence the 32Kb "Max Read". Not that it really matters, I think the transfer rate of 100MB/s makes it pretty clear that the bottleneck is the 1Gb Ethernet connection, not Starwind or the disk array.

Eventually I will get round to writing up the results of all my benchmarking properly, but I still believe the most important consideration for a SAN for virtual machine hosting is its random access transfer rate (the "real life" test in the link above). Unless you have a massive disk array this figure will be well below the max bandwidth available with 1GbE so its important to ensure that your iSCSI controller imparts as little overhead impossible, which is where software based solutions such as Starwind have a massive advantage over hardware vendors - you get to choose the hardware you are running it on and can make sure everything is up to date.

[anton (staff)]

Both. But initially you need to be 100% sure LOCAL (on the server) performance is OK. Then - check your IP SAN network. To do wire speed with TCP. Then you may start running client-side tests.