On Wed, 2005-01-19 at 16:24 -0500, James Morris wrote: > On Wed, 19 Jan 2005, Fruhwirth Clemens wrote: > > > I've rewritten some CBC code to fit the facilities I introduce in my LRW > > patch[1]. Here are the results for my P4@xxxxxx: > > > > loop-aes, CBC: ~30.5mb/s > > dm-crypt, CBC prior to my rewrite: ~23mb/s > > dm-crypt, CBC with my LRW patch: ~27mb/s > > dm-crypt, LRW with my LRW patch: ~27mb/s (slightly faster than CBC) > > > > As you can see my LRW patches (actually it's the generic scatterwalker > > which is part of the LRW patch set) halves the gap to loop-aes. > > This looks promising. I wonder if the generic scatterwalker solves the > null encryption performance problem that was reported a little while back. I've done some testing initially. The problem, at least with respect to CBC, is that the NULL cipher's block size is 1. The CBC path isn't zero originally and not even in my code (although I planned to do a zero copy version). However, my patch will surely bring some relieve, as my scatterwalk code does not give up the kmapping after every block, but wait for a page to be completely processed. So that's one kmap/kunmap instead of 4096. P.S.: The list of recipients have grown too long and topic moved a bit. I will remove a bunch of people if I post another follow-up. Please check the archive if interested and not subscribed. -- Fruhwirth Clemens <clemens@xxxxxxxxxxxxx> http://clemens.endorphin.org
Attachment:
signature.asc
Description: This is a digitally signed message part
