Soldato
- Joined
- 18 Oct 2002
- Posts
- 8,250
- Location
- The Land of Roundabouts
iperf will max out your network.
IBM Seagate Savvio 15K.2 146GB
- Thread starter Dr. Pain
- Start date
Ah had a quick look, I want to use the SAN to max out the network, hence using IOMeter to do it.
Trouble is... not many people would tailor the settings for the best case read scenario which is what I want.
I bet in the vast majority of cases people are tailoring IOMeter for a more real life workload.
I just want to max out the read capability, be it burst, sustained or whatever!
4gb/s is only 512MB/s so the SAN can easily saturate the link, I just need advice on how to get IOMeter to do it!
I think I may also Migrate the SAN stripe size, from 128K down to 64K.
All the Equallogics seem to be 64K and it does them no harm.
Associate
- Joined
- 3 Oct 2008
- Posts
- 1,890
- Location
- South London
I wouldn't bother using something like I/Ometer for that. If you want to saturate links just dump a few big ISO images across. Large contiguous files are best for generating high throughput.
Stripe size is a bit irrelevant to what you're currently trying to do as you're more than likely writing into memory. More so if you are using software initiators to connect as these often use system RAM as cache. I benched a few LUNs using M$ initiator and it used ~ 2GB RAM while doing intensive read/writes.
Stripe size is a bit irrelevant to what you're currently trying to do as you're more than likely writing into memory. More so if you are using software initiators to connect as these often use system RAM as cache. I benched a few LUNs using M$ initiator and it used ~ 2GB RAM while doing intensive read/writes.
Good point and I have seen a lot around the net on this.
Lefthand themselves recommend changing to 1, and this is the guide I followed to enable MPIO via the CLI.
Im definitely open to other suggestions, but going from 1000 to 1 really made a huge difference.
because of the switching overhead (time) you might find the performance increases if you allow it to queue some commands before it switches to the next channel otherwise you could end up with the write/read finishing before you have finished moving channel and you get no i/o for a short amount of time.
its a balancing act and I don't think there is a 1 size fits all number other than I doubt its 1000. It will porbably be a different value if you use jumbo frames for example or depending on the number of paths you have active to the array (2 or more)
I've seen docs that recommend 8 but I got better performance with 3 on my kit in my config
its a balancing act and I don't think there is a 1 size fits all number other than I doubt its 1000. It will porbably be a different value if you use jumbo frames for example or depending on the number of paths you have active to the array (2 or more)
I've seen docs that recommend 8 but I got better performance with 3 on my kit in my config