Michael van Elst a écrit : > joel.bertr...@systella.fr (=?UTF-8?Q?BERTRAND_Jo=c3=abl?=) writes: > >> fcfs seems to be a trivial FIFO queue. A have found few informations >> about priocscan strategy. But what is the difference between disksort >> and priocscan ? And between priocscan and readprio ? Does prioscan works >> in read and write and readprio only in read mode ? >
Thanks for your answer.
> fcfs is a simple FIFO
>
> disksort sorts requests by block address
>
> The queue is executed in ascending order (one-way).
> When it reaches the end, it continues at the start
> of the queue.
>
>
> readprio has two priorities
> 1. reads are queued in FIFO order
> 2. writes are sorted by block address like disksort.
>
> For reads there is a burst limit of 48 I/O request, then
> it will allow up to 16 write requests.
>
>
> priocscan has three priorities
> 1. "time critical"
> 2. "time limited" (the default)
> 3. "noncritical"
>
> the priority is chosen by the client. E.g. the filesystem
> will put synchronous metadata operations into "time critical",
> most other I/O into "time limited" and asynchronous data
> writes into "noncritical".
>
> Again each priority has a burst limit so that a queue cannot
> starve a lower priority queue.
> - "time critical" allows a burst of 4 I/Os,
> - "time limited" allows a burst of 16 I/Os,
> - "noncritical" allows a burst of 64 I/Os.
Clear.
>
>
> For stacked device drivers (like dk,cgd,ccd or vnd), it's usually
> better to use simple FIFO on the upper layers and to choose the
> buffer strategy only for the lowest layer that actually talks
> to hardware.
I don't understand. For example, my rootfs is on /dev/raid0a.
legendre# dkctl raid0a
strategy:
/dev/rraid0a: priocscan
getcache:
/dev/rraid0a: read cache enabled
/dev/rraid0a: write-back cache enabled
/dev/rraid0a: read cache enable is not changeable
/dev/rraid0a: write cache enable is changeable
/dev/rraid0a: cache parameters are not savable
/dev/rraid0a: cache Force Unit Access (FUA) supported
/dev/rraid0a: cache Disable Page Out (DPO) not supported
listwedges:
/dev/rraid0a: no wedges configured
legendre# dkctl raid0
strategy:
/dev/rraid0d: priocscan
...
If I change strategy on raid0 device, all slices are modified with the
same strategy. And I cannot modify ccd0 strategy (nor wedges strategies
on this ccd0 device) :
legendre# dkctl ccd0
strategy:
dkctl: /dev/rccd0: DIOCGSTRATEGY: Inappropriate ioctl for device
legendre# dkctl ccd0a
strategy:
dkctl: /dev/rccd0a: DIOCGSTRATEGY: Inappropriate ioctl for device
legendre# dkctl dk3
strategy:
dkctl: /dev/rdk3: DIOCGSTRATEGY: Inappropriate ioctl for device
legendre#
> On the other hand, the upper layers rarely queue anything, so
> the difference is just how much CPU time is wasted in processing
> an expensive strategy on multiple layers.
>
>
> Modern disks queue many requests themselves, the strategy
> used by the kernel has little meaning then. Just like above,
> the stragey done by the lowest layer counts.
That being said, if I understand, disksort or readprio could be more
efficient than priocscan.
Regards,
JB
signature.asc
Description: OpenPGP digital signature
