On Wed, Oct 25 2017 at 4:58am -0400, Yael Chemla <yael.chemla@xxxxxxxxxxxx> wrote: > Current implementation makes multiple crypto API calls for a single > block, forcing underlying crypto tfm implementation to "stop & go", > leading to under utilization of CPU (e.g. SIMD state saves) or HW > engines. To fix it unify calls to crypto init/update/final into a digest > call with a single sg which contains multiple buffers. > > This also opens the way later on to parallelizing the work on different > blocks. > > Tested on physical Arm 32 and x86_64 and virtual ARM 64. > > Performance numbers were not changed. OK, but what about CPU utilization? Was it reduced? Thanks, Mike -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
