From mboxrd@z Thu Jan 1 00:00:00 1970 From: Vincent JARDIN Subject: Re: Scenario of dpdk-iokit //Re: Fwd: dpdk-iokit: Turn dpdk into the IO field. Date: Tue, 19 Nov 2013 11:29:23 +0100 Message-ID: <528B3D83.6040900@6wind.com> References: , , <521D9455-820A-4F8F-A7E7-8380EE1415DF@intel.com> <877C1F8553E92F43898365570816082F1ADE612D@IRSMSX103.ger.corp.intel.com> <47C784D99F4D124BB2973FE5C27BDCF12FB36570@FMSMSX109.amr.corp.intel.com> <7F974523047D074590B0CBEB8DDF43A70E516CFB@FMSMSX104.amr.corp.intel.com> <47C784D99F4D124BB2973FE5C27BDCF12FB374A9@FMSMSX109.amr.corp.intel.com> <9F232B325D833F4494EFCD5B9D49F6C719D5CFC8@IRSMSX103.ger.corp.intel.com> <00B2C372B6E6DA4FB04934511BF0564A3DE6AEC9@FMSMSX114.amr.corp.intel.com> <528B3346.3050100@6wind.com> Mime-Version: 1.0 Content-Type: text/plain; charset=windows-1252; format=flowed Content-Transfer-Encoding: quoted-printable Cc: dev-VfR2kkLFssw@public.gmane.org To: March Jane Return-path: In-Reply-To: List-Id: patches and discussions about DPDK List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces-VfR2kkLFssw@public.gmane.org Sender: "dev" so on the list if you prefer to do so ;) did you try a "perf top" on=20 your system under load? On 19/11/2013 10:56, March Jane wrote: > > Gratitude to reply. > =09 > FCP and Full-offload iSCSI will be nice, for example Qlogic 16Gbps FCP= . > =09 > But IO scenario is very different from IP scenario as my understanding= . > > March=09 > > BTW: Do you have people at Bay Area? > > > PS: > Already finished header file for dpdk-iokit/libiokit_sctgt > > =09 > > > > > =09 > =09 > > > > > > > > On Nov 19, 2013, at 5:45 PM, Vincent JARDIN = wrote: > >> (off list since it coul become a troll ;) ) >> >> At 6WIND, we developed a librte_crypto as a generic framework to manag= e any crypto framework: >> - Intel's QuickAssist >> - Cavium' Nitrox II >> - AES/NI SW crypto >> >> it was required in order to manage high rate of PCI IOs for Cryptos (I= Psec, SSL, etc...) >> >> So, after the librte_pmd_mlx4, Virtio, Vmxnet3, that's more than 6 ult= ra low latency/very efficient drivers that we added into the DPDK and pro= moted. >> >> Which storage drivers would you foresee first to be run in userland? >> >> Best regards, >> >> >> On 19/11/2013 02:22, March Jane wrote: >>> >>> =97 Nowadays dilemmas in enterprise-class storage system & motivati= on -- >>> >>> The model of most storage system is =91Front-end cards + CPU += back-end magnetic media hard drive=92, in such system, the hard drive is= very slow in terms of its capability in random accessing - roughly 200IO= per second, in contrast, the processor is very fast, therefore many soft= ware stack can tolerate waste of CPU cycle for exploiting capabilities of= hard drive. >>> =09 >>> In addition, regularly software for SAN is developed under kernel sp= ace in order to achieve low latency - few milliseconds per request, and I= O pressure in scenario of heavy workload. Usually software in user-space = with POSIX is slow. However kernel-space developing is nightmare for engi= neers, even we have approaches to simulate these code under user-space. >>> >>> However, so far flash is coming to popular, the gap between CPU = and media is overturned, a single flash card can be easy to reach 1M iops= , if plug 10 such cards inside a server, the processor is hard to back. T= hus, today=92s challenge in flash storage is to exploit the capability of= processor, ironically. In addition to this purpose, move to user-space = is also a motivation, if move to user-space is OS-bypass rather than movi= ng to POSIX. >>> >>> >>> Best >>> >>> - March >>> >>> >>> >