[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [dm-devel] dm-crypt parallelization patches




On Tue, 3 Sep 2013, Andi Kleen wrote:

> On Tue, Sep 03, 2013 at 03:06:03PM -0400, Mikulas Patocka wrote:
> > Hi Andi
> > 
> > You wrote the original dm-crypt paralllelization patch (commit 
> > c029772125594e31eb1a5ad9e0913724ed9891f2). This patch improves 
> > parallelization, it makes data being encrypted on the same CPU that 
> > submitted the request.
> 
> The motivation for my patches was to make large systems with many
> cores scale.
> 
> I think the motivation for yours is to make small systems without
> hardware crypto offload slightly faster.
> 
> That's fine (although I suspect it's more and more obscure), as long as
> you don't impact the large systems scalability.
> 
> Can you do some tests with a larger system with very parallel IO
> (many CPUs submitting)  and see if the IO rate is equivalent?
> 
> If that's given the patches are ok for me.
> 
> -Andi

The patches mainly help with sequential I/O (sequential I/O is normally 
submitted from a few CPUs, so existing dm-crypt doesn't parallelize 
right). There is a decrease of performance when 12 threads are doing 
writes on ramdisk:

Sequential read of two SCSI disks in RAID-0 (throughput):
	raw read rate: 262MB/s stdev 2.9
	existing dm-crypt: 111MB/s stdev 19.4 (it varies greatly)
	new dm-crypt: 254MB/s stdev 1.5

Sequential read of ramdisk (throughput):
	raw read rate: 908MB/s stdev 7.7
	existing dm-crypt: 133MB/s stdev 0.8
	my new dm-crypt: 475MB/s stdev 6

fio --rw=randread --bs=4k --direct=1
Random read (4k block) on ramdisk from 12 threads (time in seconds):
	raw device: 2.32s stdev 0.02
	existing dm-crypt: 17.95s stdev 3.40
	new dm-crypt: 15.72s stdev 1.86

fio --rw=randwrite --bs=4k --direct=1
Random write (4k block) on ramdisk from 12 threads (time in seconds):
	raw device 3.91s stdev 0.01
	existing dm-crypt: 21.16s stdev 3.67
	new dm-crypt: 27.42s stdev 0.27

fio --rw=randrw --bs=4k --direct=1
Random read+write (4k block) on ramdisk from 12 threads (time in seconds):
	raw device: 3.09s stdev 0.01
	existing dm-crypt: 20.08s stdev 3.48
	new dm-crypt: 24.78s stdev 0.13

Tests were done on two six-core Istanbul-class Opterons.

I don't have a computer with AES-NI, so I can't test that.

Mikulas


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]