Single-threaded sequential write performance issue

[diamondlod]

I'm using a 3PAR 7450 with mostly MFC, and, as expected, we're having issues with an application that needs to sustain a high MB/sec rating in 40k blocks on a single thread.

Can anyone suggest a config that will get us the best performance under these circumstances?

Thanks,

Di

[Cleanur]

With a single threaded workload I'd start by turning off interrupt coalescing on the host ports.That way there'll be no I/O wait while the port buffer fills, instead each incoming I/O will generate it's own interrupt and be processed immediately.

As you've already implied the arrays ability to perform is highly dependent on the applications ability to deliver sufficient sustained workload, so you may also have to take a look back up the I/O stack for further improvements.

[diamondlod]

I was able to disable the interrupt coalescing, and it definitely helped, but not nearly enough. What we really need is for a single thread to take more bandwidth, it appears.

If anyone has other suggestions, I'm all ears.

Thanks

[Cleanur]

You could try adjusting the queue depth at the host. The problem is you need to load up the array with outstanding IO in order to keep the data flowing. With a single threaded workload that's typically not possible against a multi-core equipped system, it sounds like you application in it's current state just isn't capable of driving the I/O to it's max.

[diamondlod]

Actually those two steps combined helped dramatically! We're sustaining 330MB/sec at 64k!

Thank you!