[mythtv-users] Software RAID and ionice

[David Rees]

On 5/22/07, Richard Freeman <r-mythtv at thefreemanclan.net> wrote:
> The raid5 works just fine for mythbackend when the system isn't busy
> writing to the same disk.

Which indicates a bottleneck somewhere. ext3 is notorious for blocking
writes when the journal is flushing which kills performance. Did you
create the filesystem with the stride option? How big are the stripes
in your raid array? How many disks and how fast are they?

As I said before, for some workloads which involve small writes, RAID5
can be significantly slower than a single disk. When ext3 flushes the
journal to disk and it is writing out inode information, a lot of
small writes scattered all over the disk will end up blocking all
writes to the array for some period of time. ionice or not, your
application is going to get stuck if it is waiting for data to be
written to disk. jfs also suffers from this from some degree. xfs is
much better about laying out data on the disk and doesn't suffer from
this.

This is one reason I wish mythtv didn't sync all writes to disk and
instead let the OS handle it on it's own since then the OS could queue
up the write and and write out the data when it can.

> As far as xfs goes - I agree that it tends to perform better, but the
> only thing that concerns me about xfs is that it wipes out any file that
> is open if it isn't cleanly unmounted.  I used to use xfs and migrated
> to ext3 for precisely that reason.

The only issue xfs has that I knew about was if the file hadn't synced
to disk, it would be possible for the file to contain null data.

Perhaps you can partition your system to use xfs for the myth storage
and ext3 for the rest which would solve the problem. Even if you used
ext3 on both partitions that would likely improve the situation.

-Dave