Page 1 of 1

How to improve HA replication (FULL Sync)

Posted: Tue Dec 21, 2010 11:46 am
by patrick1140
Hello,

On a 1GB link (i've one dedicated link by HA volume) if i launch a full sync, the network usage for this network card is limited to 50% . Any solution to change some parameters into starwind to use 100% ?

Re: How to improve HA replication (FULL Sync)

Posted: Tue Dec 21, 2010 2:28 pm
by anton (staff)
50% network usage is a normal value for our current implementation. There's no way to see 100% with any tuning settings. You can use cross-link with higher bandwidth (10 GbE or trunk * 1 GbE).
patrick1140 wrote:Hello,

On a 1GB link (i've one dedicated link by HA volume) if i launch a full sync, the network usage for this network card is limited to 50% . Any solution to change some parameters into starwind to use 100% ?

Re: How to improve HA replication (FULL Sync)

Posted: Mon Dec 27, 2010 8:55 pm
by clayton@mcc911.org
Patrick

assuming you are using your network adapter in fullduplex ie with a switch or crossover cable.
in windows the network meter adds your send and receave together, so if you have a 1Gbps card your total max bandwith is 2Gbps, 1Gbps sending and 1Gbps receaving, and in that scale 50% would be the max if it is a one way with low overhead, what you need are numbers it would be much more handly to know your Mbps , get you a network meter for testing something like DUmeter.

Re: How to improve HA replication (FULL Sync)

Posted: Tue Dec 28, 2010 5:09 pm
by Max (staff)
Clayton,
thanks for the useful info, we will also recommend it, in some cases people rely only on the Windows internal tools

Re: How to improve HA replication (FULL Sync)

Posted: Tue Dec 28, 2010 7:04 pm
by anton (staff)
At this moment during performing re-sync sequence one node holds both so-called "data valid" token and "data pusher" token as well. In the other words - it's one node moving sync data over the network in only one direction... It's possible however to skip using "data pusher" token and slide away from master-slave re-sync processing, so both nodes will cross-send at the same time. Doing this we'll fully utilize full-duplex 1 GbE and 10 GbE wiring nature and will achieve both 100% network utilization and cut sync time in a half. You'll see this in one of the upcoming StarWind editions. Thanks!

Re: How to improve HA replication (FULL Sync)

Posted: Tue Dec 28, 2010 7:43 pm
by clayton@mcc911.org
Oh I See!!!
So for example if you have two starwind boxes in HA, and an iSCSI Client is going to write 1000Mb of data, Using Mutipathing
the client would send 500Mb to Node1 and 500Mb to Node 2 by alternating every other packet. So each node would have both 500Mb of Data to send sync and 500Mb of Data to receive sync, so it could leverage the fullduplex connection with equal send and receving traffic.

This is very cool and smart , Good Job
when will get get to see this working? Soon I hope
Clayton

Re: How to improve HA replication (FULL Sync)

Posted: Wed Dec 29, 2010 1:36 pm
by Aitor_Ibarra
Hi Clayton,

Starwind already does that (multipathing active/active) - you just need the target & partner to be in sync. I'm sure that there are more performance improvements to come beyond the improvements to full sync that Anton was mentioning. Having said that, I was quite impressed recently with the full sync performance when I had recently to do a full sync of a 2TB target which was on slow 5.4K rpm sata drives - the bottleneck was definitely the drives!

Although Starwind can now enable a target for initiator connections while it is undergoing a full sync, I wouldn't recommend this if the workload of the target is high, as the full sync really taxes the drives. Better to keep i/o to a minimum or even suspend it while a full sync is taking place. It would be nice to be able to limit the bandwidth starwind uses for full sync in these situations - full sync would take longer, of course, but then anything using the target won't slow to a crawl.

cheers,

Aitor

Re: How to improve HA replication (FULL Sync)

Posted: Thu Dec 30, 2010 9:35 am
by anton (staff)
Guys, don't mix pure Active-Active "normal" working scenario and muli-node activity during re-sync process. They are quite different and totally independent from each other things - Aitor is 100% correct on this ))