Intel FakeRaid and image sequences

Hi,

I've been having problems with the write performance using intel's onboard raid5 implementation. The performance of the raid seems to drop drastically when writing image sequences vs single files.

The array consists of 3 3Tb Seagate 7200RPM drives.

Does anyone have any advice regarding improving the performance of Intel's fake raid?

Cheers,

Andrew

[Joshua Helling]

Write speed on RAID 5 is always slower than it's read due to the Parity calculation. For single long files this tends to be less invasive, but with several files (read:sequenced images) this will be worse because parity is calculated for EACH file.

So, to make some calculations here. A three drive RAID 5 will have a throughput of around 120 MB write and 150 read....SATA3 Drives and Controller may make that faster, but I'm assuming SATA2 in this case.

So these are also kinda optimal numbers. Format will be an important factor here. For example 1080i59.94 10Bit YUV uncompressed needs 156MB/second. Which is over your realistically consistant throughput. If you're doing our DPX it's RGB it's a bit bigger (due to being 4:4:4 and not 4:2:2). As the resolution goes down so does the requirement, so you might be using some other format like 720p or 486i.

So I guess the first thing I'd do is find out the data rate of your sequence and then run the disk speed test on your RAID (keeping in mind what I said above about a single file requires less overhead than a sequence).

Hey Joshua,

Thanks for your reply.

They are indeed sata2 drives connected to a sata 2 chip set (I'm saving the remaining 2 sata3 ports on the mobo for an ssd carrier when I get my BMCC).

After doing further testing, I do indeed have pretty poor write performance (avg about 70Mb/sec) on the array but I can improve this using the Volume Writeback Cache. That improves performance on the array to 350ish Mb/sec read and 250 Write.

What then seems to happen is that it will run at 50+ frames a second when set to render at maximum fps for around 20 second and then after that, with disk throughput at 200Mb/s. After, something seems to happen to the writeback cache and performance drops to 50Mb/s, lower than what the array should be capibile of without the writeback cache.

In another post, you mentioned that you have this Mobo in house for testing. I was wondering if you noticed the same writeback cache behavior?

[Joshua Helling]

Well, the writeback cache is kinda like a stop gap. It will improve burst performance, but in the long haul you will suffer as you've seen.

The best way to improve performance is to add more drives to the array set. Or you could drop to RAID0 (you'll loose the redundancy) but that still might not provide enough throughput consistently.

When we test these rigs we either test them on large externally attached arrays, or 4 drive internal RAID0. We don't thoroughly test the SATA connections as that's not what we are testing. We make sure that our boards get enough throughput to support uncompressed capture/playback at the highest device supported resolutions.