From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail.droids-corp.org (zoll.droids-corp.org [94.23.50.67]) by dpdk.org (Postfix) with ESMTP id AF63E5A6C for ; Mon, 15 Jun 2015 16:30:02 +0200 (CEST) Received: from was59-1-82-226-113-214.fbx.proxad.net ([82.226.113.214] helo=[192.168.0.10]) by mail.droids-corp.org with esmtpsa (TLS1.2:DHE_RSA_AES_128_CBC_SHA1:128) (Exim 4.80) (envelope-from ) id 1Z4VTH-000873-Md; Mon, 15 Jun 2015 16:34:47 +0200 Message-ID: <557EE1A0.609@6wind.com> Date: Mon, 15 Jun 2015 16:30:56 +0200 From: Olivier MATZ User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:31.0) Gecko/20100101 Icedove/31.6.0 MIME-Version: 1.0 To: Bruce Richardson References: <87110795-201A-4A1E-A4CC-A778AA7C8218@cisco.com> <557ED116.7040508@6wind.com> <20150615134409.GA7500@bricha3-MOBL3> <2601191342CEEE43887BDE71AB97725836A0A838@irsmsx105.ger.corp.intel.com> <557EDB91.9010503@6wind.com> <20150615141258.GA580@bricha3-MOBL3> In-Reply-To: <20150615141258.GA580@bricha3-MOBL3> Content-Type: text/plain; charset=windows-1252 Content-Transfer-Encoding: 7bit Cc: "dev@dpdk.org" , "Damjan Marion \(damarion\)" Subject: Re: [dpdk-dev] rte_mbuf.next in 2nd cacheline X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.15 Precedence: list List-Id: patches and discussions about DPDK List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Mon, 15 Jun 2015 14:30:02 -0000 On 06/15/2015 04:12 PM, Bruce Richardson wrote: > On Mon, Jun 15, 2015 at 04:05:05PM +0200, Olivier MATZ wrote: >> Hi, >> >> On 06/15/2015 03:54 PM, Ananyev, Konstantin wrote: >>> >>> >>>> -----Original Message----- >>>> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Bruce Richardson >>>> Sent: Monday, June 15, 2015 2:44 PM >>>> To: Olivier MATZ >>>> Cc: dev@dpdk.org; Damjan Marion (damarion) >>>> Subject: Re: [dpdk-dev] rte_mbuf.next in 2nd cacheline >>>> >>>> On Mon, Jun 15, 2015 at 03:20:22PM +0200, Olivier MATZ wrote: >>>>> Hi Damjan, >>>>> >>>>> On 06/10/2015 11:47 PM, Damjan Marion (damarion) wrote: >>>>>> >>>>>> Hi, >>>>>> >>>>>> We noticed 7% performance improvement by simply moving rte_mbuf.next field to the 1st cache line. >>>>>> >>>>>> Currently, it falls under /* second cache line - fields only used in slow path or on TX */ >>>>>> but it is actually used at several places in rx fast path. (e.g.: i40e_rx_alloc_bufs() is setting that field to NULL). >>>>>> >>>>>> Is there anything we can do here (stop using next field, or move it to 1st cache line)? >>>>> >>>>> Agree, this is also something I noticed, see: >>>>> http://dpdk.org/ml/archives/dev/2015-February/014400.html >>>>> >>>>> I did not have the time to do performance testing, but it's something >>>>> I'd like to do as soon as I can. I don't see any obvious reason not to >>>>> do it. >>>>> >>>>> It seems we currently just have enough room to do it (8 bytes are >>>>> remaining in the first cache line when compiled in 64 bits). >>>> >>>> This, to me, is the obvious reason not to do it! It prevents us from taking in >>>> any other offload fields in the RX fast-path into the mbuf. >>>> >>>> That being said, I can see why we might want to look to move it - but from the >>>> work done in the ixgbe driver, I'd be hopeful we can get as much performance with >>>> it on the second cache line for most cases, through judicious use of prefetching, >>>> or otherwise. >>>> >>>> It took a lot of work and investigation to get free space in the mbuf - especially >>>> in cache line 0, and I'd like to avoid just filling the cache line up again as >>>> long as we possibly can! >>> >>> Yep, agree with Bruce here. >>> Plus, with packet_type going to be 4B and vlan_tci_outer, >>> we just don't have 8 free bytes at the first cache line any more. >> >> I don't understand why m->next would not be a better candidate than >> rx offload fields to be in the first cache line. For instance, m->next >> is mandatory and must be initialized when allocating a mbuf (to be >> compared with m->seqn for instance, which is also in the first cache >> line). So if we want to do some room in the first cache line, I >> think we can. >> >> To me, the only reason for not doing it now is because we don't >> have a full performance test report (several use-cases, drivers, ...) >> that shows it's better. >> > Because the "next" field is not mandatory to be set on initialization. It can > instead be set only when needed, and cleared on free if it is used. > > The next pointers always start out as NULL when the mbuf pool is created. The > only time it is set to non-NULL is when we have chained mbufs. If we never have > any chained mbufs, we never need to touch the next field, or even read it - since > we have the num-segments count in the first cache line. If we do have a multi-segment > mbuf, it's likely to be a big packet, so we have more processing time available > and we can then take the hit of setting the next pointer. Whenever we go to > free that mbuf for that packet, the code to do the freeing obviously needs to > read the next pointer so as to free all the buffers in the chain, and so it can > also reset the next pointer to NULL when doing so. > > In this way, we can ensure that the next pointer on cache line 1 is not a problem > in our fast path. This is a good idea, but looking at the drivers, it seems that today they all set m->next to NULL in the rx function. What you are suggesting is to remove all of them, and document somewhere that all mbufs in a pool are supposed to have their m->next set to NULL, correct? I think what you are describing could also apply to reference counter (set to 1 by default), right? Olivier > > /Bruce >