[ Thursday, September 16, 1999 ] [EMAIL PROTECTED] wrote:
> i'll have a look at RAID1 read balancing. I once ensured we read better
> than single-disk, but we might have lost this property meanwhile ...
Lingering question:
- Can the 128 sector count for switching be changed safely?
if so, I'd love to see something in /proc I could echo a new number into
to tune the disk switching to my particular access patterns...
Whether drives can scatter/gather well, how large their caches are,
how fast they seek, how aggressively they read ahead, how many drives
in the mirror, etc, etc... they would all (among many other variables)
affect the "optimal" setting for number of sectors to hit a drive up for
until switching... the raid10 case, along with s/w raid over h/w raid,
could further (hugely) exacerbate the problem...
Perhaps with large enough block sizes (does the blocks that
raid1_make_request() deal with exist as 4K in the raid1 case? 8K? larger?)
I'd love to see the possibility of splitting a block request into 2 block
requests to the drives... so an 8K block request of raid1_make_request()
creates two 4K requests to each drive... then we could toss (I believe)
the other cruft for read-balancing *shrug*... at this point I'd take
being able to drop the 128 down and see if that improves balancing.
James
--
Miscellaneous Engineer --- IBM Netfinity Performance Development
