More observations on my issue: > Did you dd against the raw device, the block device, or against the file > system on top of the raid? I first tried against the fs. I now added a scratch partition and tried the devices: I get 40 to 50MB/s on the raw device and a whopping 0.57MB/s on the block device! Nevertheless, when writing to the block dev, sysstat shows the raid being 100% busy and each of the discs being 50-60% busy. Moreover, when writing to the block dev, dd's WCHAN is vnode!
> Did you try to match up the raid data stripe size with the filesystem > block size? No, I didn't. But it is aligned by chance. > What kind of performance do you get if you do the following: > dd if=/dev/zero of=/dev/rraid0d bs=<data_stripe_size> count=100 On a scratch raid2h, some 0,57MB/s. > Do you get reasonable write performance for each of the component disk in > the raid? I.e. is there any possibility of partially failing disk? Yes. 21MB/s for each for them. > I noticed you tried raid1, which only takes 2 disk, so is the 3rd disk okay? The RAID1s are across three discs, too. Two more random observations: I continuously see ~3763 Interrupts on ioapic0 pin4, yes dmesg doesn't show anything for that pin. Nevertheless, there is ~0% CPU time spent in interrupt mode. When un-tar-ing src.tgz, the raid (plus it's components) are busy for minutes after the command finishes. Looks like there's something seriously wrong here, especially the raw dev being a hundred times faster than the block dev.
