On Tuesday August 24, ptb@xxxxxxxxxx wrote:
> In raid1 one could read from both (or more drives) and discard all read
> returns except the first. This should be fractionally faster than
> reading from one drive alone in most circumstances.
>
> However, one can also split the reads so that if say 128 requests are
> queued, then the lowest 64 are read from one drive and the highest 64
> from another. With luck, that would amount to doubling the read rate,
> as there would be little overlap on what is read from the two drives,
> and requests are issued asyncronously.
>
> Curiously, that is about what the current code does (by accident, and
> slightly wrongly), but it stops doing it in exactly the circumstances
> when it would do good - during a linear set of contiguous reads. In
> that case one should issue 50% of each batch of requests given to the
> request function to one drive, and the other 50% as far away as possible
> to the other drive. The current code does no switch half way through
> each batch of requests (where it would do good), but instead after
> counting through exactly 128 requests, which is a pretty random point
> in a batch.
>
> If the reads come from all over the shop, then one still gains from
> doing half on one drive and half on the other, so long as one does
> both halves at the same time and one arranges that the readaheads
> do not overlap the reads from the other drives (it's easier to say
> that they don't when reads are a contiguous linear sequence, which is
> why I said that that's the case when one should be sharing reads
> between two drives).
Sounds like "plugging" might be the thing to do.
When a read request arrives, just stick it on a queue, counting
the number of sectors.
When an "unplug" call is made:
Sort the queue
Submit half the sectors to one drive, and the other half
to the other drive.
Further refinements:
Not "half" but "1/N" if there are "N" drives.
Keep track of number of outstanding reads per drive, and
submit fewer requests to busier drives.
Does that sound like what you want?
NeilBrown
-
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
