On Mon, 2014-07-14 at 20:17 +0200, Peter Zijlstra wrote: > On Mon, Jul 14, 2014 at 10:05:34AM -0700, Tim Chen wrote: > > I was trying to explain why the algorithm is implemented this way > > because of its batching nature. > > > > There is a whole class of async algorithm that can provide > > substantial speedup by doing batch processing and uses workqueue. > > The multi-buffer sha1 version has 2.2x speedup over existing > > AVX2 version, and can have even more speedup when AVX3 > > comes round. Workqueue is a natural way to implement > > this. I don't think a throughput speedup of 2.2x is "crap". > > > > We are not inventing anything new, but ask for a > > very simple helper function to know if there's something else > > running on our cpu to help us make a better decision > > of whether we should flush the batched jobs immediately. > > > > And also asynchronous crypto interface is already used substantially > > in crypto and has a well established infrastructure. > > The crap I was talking about is that there's a metric ton of 'async' > interfaces all different. Async interfaces when used appropriately, actually speed things up substantially for crypto. We actually have a case with ecyrptfs not using the async crypto interface, causing cpu to stall and slowing things down substantially with AES-NI. And async interface with workqueue speed things up (30% to 35% on encryption with SSD). http://marc.info/?l=ecryptfs-users&m=136520541407248 http://www.spinics.net/lists/ecryptfs/msg00228.html > > Your multi-buffer thing isn't generic either, it seems lmiited to sha1. We actually have many other multi-buffer crypto algorithms already published for encryption and other IPSec usages. So multi-buffer algorithm is not just limited to SHA1. We hope to port those to the kernel crypto library eventually. http://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-multi-buffer-ipsec-implementations-ia-processors-paper.pdf > It does not reuse padata, padata tries to speed things up by parallelizing jobs to *multiple* cpus. Whereas multi-buffer tries to speed things up by speeding things up by using multiple data lanes in SIMD register in a *single* cpu. These two usages are complementary but not the same. > it does not extend workqueues, Why do I need to extend workqueues if the existing ones already meet my needs? > it does not > remove the btrfs nonsense, Not much I can do about btrfs as I don't understand the issues there. > it adds yet anotehr thing. Thanks. Tim -- To unsubscribe from this list: send the line "unsubscribe linux-crypto" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html