From mboxrd@z Thu Jan 1 00:00:00 1970 From: Thomas Monjalon Subject: Re: [PATCH v4] crypto/aesni_gcm: migration from MB library to ISA-L Date: Thu, 05 Jan 2017 16:12:26 +0100 Message-ID: <17462396.FW1IWXuNnf@xps13> References: <1483448523-150510-1-git-send-email-piotrx.t.azarewicz@intel.com> <1483624286-92526-1-git-send-email-piotrx.t.azarewicz@intel.com> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7Bit Cc: dev@dpdk.org, pablo.de.lara.guarch@intel.com To: Piotr Azarewicz Return-path: Received: from mail-wm0-f52.google.com (mail-wm0-f52.google.com [74.125.82.52]) by dpdk.org (Postfix) with ESMTP id 6CAFF5A44 for ; Thu, 5 Jan 2017 16:12:28 +0100 (CET) Received: by mail-wm0-f52.google.com with SMTP id c85so249753171wmi.1 for ; Thu, 05 Jan 2017 07:12:28 -0800 (PST) In-Reply-To: <1483624286-92526-1-git-send-email-piotrx.t.azarewicz@intel.com> List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org Sender: "dev" 2017-01-05 14:51, Piotr Azarewicz: > Current Cryptodev AES-NI GCM PMD is implemented using Multi Buffer > Crypto library.This patch reimplement the device using ISA-L Crypto > library: https://github.com/01org/isa-l_crypto. > > The migration entailed the following additional support for: > * GMAC algorithm. > * 256-bit cipher key. > * Session-less mode. > * Out-of place processing > * Scatter-gatter support for chained mbufs (only out-of place and > destination mbuf must be contiguous) > > Verified current unit tests and added new unit tests to verify new > functionalities. > > PERFORMANCE COMPARISON > ---------------------- > Comparison the new and old implementation is made by running app/test > and calling cryptodev_aesni_gcm_perftest. > As we may see below, the new implementation has small performance drop > when buffer size is above 64B. I am a bit surprised that you change for lower performance. I understand that it brings new features. Is it possible to add such features without performance impact? > Old implementation with MB library: > Cipher algo: AES_GCM Cipher hash: AES_GCM ciphr key: 128b burst size: 32 > Buffer Size(B) OPS(M) Throughput(Gbps) Retries EmptyPolls > 64 4.57 2.34 0 0 > 128 4.28 4.39 0 0 > 256 2.76 5.66 0 0 > 512 1.60 6.56 0 0 > 1024 0.90 7.34 0 0 > 1536 0.62 7.66 0 0 > 2048 0.48 7.84 0 0 > > New implementation with ISA-L library: > Cipher algo: AES_GCM Cipher hash: AES_GCM ciphr key: 128b burst size: 32 > Buffer Size(B) OPS(M) Throughput(Gbps) Retries EmptyPolls > 64 4.62 2.37 0 0 > 128 4.06 4.16 0 0 > 256 2.65 5.44 0 0 > 512 1.57 6.45 0 0 > 1024 0.89 7.26 0 0 > 1536 0.62 7.58 0 0 > 2048 0.47 7.77 0 0 >