> > On Sun, Dec 7, 2014 at 4:23 PM, Weiny, Ira <ira.we...@intel.com> wrote: > > I don't think this is a bad idea, however, the complication is determining > > the > kmem_cache object size in that case. How about we limit this value to < 2K > until such time as there is a need for larger support? > > Can we just get rid of all the kmem_caches used in the MAD code? I don't > think any of this is so performance-critical that we couldn't just use > kmalloc... >
On an SM node the MAD code is _very_ performance-critical. I don’t think we should give that up. Another alternative is to have a cache per device based on the size reported. -- Ira N�����r��y����b�X��ǧv�^�){.n�+����{��ٚ�{ay�ʇڙ�,j��f���h���z��w��� ���j:+v���w�j�m��������zZ+�����ݢj"��!�i