Date: Mon, 9 May 2016 22:31:37 +0300 From: Slawa Olhovchenkov <slw@zxy.spb.ru> To: John Baldwin <jhb@freebsd.org> Cc: src-committers@freebsd.org, svn-src-head@freebsd.org, svn-src-all@freebsd.org Subject: Re: svn commit: r299210 - in head/sys/dev/cxgbe: . tom Message-ID: <20160509193137.GH1447@zxy.spb.ru> In-Reply-To: <2354770.oBAqoHF8jb@ralph.baldwin.cx> References: <201605070033.u470XZCs075568@repo.freebsd.org> <3833131.rOKpC7i1Gu@ralph.baldwin.cx> <20160509185719.GG1447@zxy.spb.ru> <2354770.oBAqoHF8jb@ralph.baldwin.cx>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, May 09, 2016 at 12:03:22PM -0700, John Baldwin wrote: > On Monday, May 09, 2016 09:57:19 PM Slawa Olhovchenkov wrote: > > On Mon, May 09, 2016 at 10:49:30AM -0700, John Baldwin wrote: > > > > > On Saturday, May 07, 2016 04:44:51 PM Slawa Olhovchenkov wrote: > > > > On Fri, May 06, 2016 at 05:52:15PM -0700, John Baldwin wrote: > > > > > > > > > On Saturday, May 07, 2016 12:33:35 AM John Baldwin wrote: > > > > > > Author: jhb > > > > > > Date: Sat May 7 00:33:35 2016 > > > > > > New Revision: 299210 > > > > > > URL: https://svnweb.freebsd.org/changeset/base/299210 > > > > > > > > > > > > Log: > > > > > > Use DDP to implement zerocopy TCP receive with aio_read(). > > > > > > > > > > > > Chelsio's TCP offload engine supports direct DMA of received TCP payload > > > > > > into wired user buffers. This feature is known as Direct-Data Placement. > > > > > > However, to scale well the adapter needs to prepare buffers for DDP > > > > > > before data arrives. aio_read() is more amenable to this requirement than > > > > > > read() as applications often call read() only after data is available in > > > > > > the socket buffer. > > > > > > > > > > > > When DDP is enabled, TOE sockets use the recently added pru_aio_queue > > > > > > protocol hook to claim aio_read(2) requests instead of letting them use > > > > > > the default AIO socket logic. The DDP feature supports scheduling DMA > > > > > > to two buffers at a time so that the second buffer is ready for use > > > > > > after the first buffer is filled. The aio/DDP code optimizes the case > > > > > > of an application ping-ponging between two buffers (similar to the > > > > > > zero-copy bpf(4) code) by keeping the two most recently used AIO buffers > > > > > > wired. If a buffer is reused, the aio/DDP code is able to reuse the > > > > > > vm_page_t array as well as page pod mappings (a kind of MMU mapping the > > > > > > Chelsio NIC uses to describe user buffers). The generation of the > > > > > > vmspace of the calling process is used in conjunction with the user > > > > > > buffer's address and length to determine if a user buffer matches a > > > > > > previously used buffer. If an application queues a buffer for AIO that > > > > > > does not match a previously used buffer then the least recently used > > > > > > buffer is unwired before the new buffer is wired. This ensures that no > > > > > > more than two user buffers per socket are ever wired. > > > > > > > > > > > > Note that this feature is best suited to applications sending a steady > > > > > > stream of data vs short bursts of traffic. > > > > > > > > > > > > Discussed with: np > > > > > > Relnotes: yes > > > > > > Sponsored by: Chelsio Communications > > > > > > > > > > The primary tool I used for evaluating performance was netperf's TCP stream > > > > > test. It is a best case for this (constant stream of traffic), but that is > > > > > also the intended use case for this feature. > > > > > > > > > > Using 2 64K buffers in a ping-pong via aio_read() to receive a 40Gbps stream > > > > > used about about two full CPUs (~190% CPU usage) on a single-package > > > > > Intel E5-1620 v3 @ 3.50GHz with the stock TCP stack. Enabling TOE brings the > > > > > usage down to about 110% CPU. With DDP, the usage is around 30% of a single > > > > > CPU. With two 1MB buffers the the stock and TOE numbers are about the same, > > > > > but the DDP usage is about 5% of single CPU. > > > > > > > > > > Note that these numbers are with aio_read(). read() fares a bit better (180% > > > > > for stock and 70% for TOE). Before the AIO rework, trying to use aio_read() > > > > > with two buffers in a ping-pong used twice as much CPU as bare read(), but > > > > > aio_read() in general is now fairly comparable to read() at least in terms of > > > > > CPU overhead. > > > > > > > > Can be this impovement of nfsclient and etc? > > > > > > The NFS client is implemented in the kernel (and doesn't use the AIO > > > interfaces), so that would be a bit trickier to manage. OTOH, this could be > > > useful for something like rsync if that had an opton to use aio_read(). > > > > May be possible by some additional create some general API > > for using inside kernel for nfsclient/nfsd/iscsi initiator/target/etc? > > Automatic using, in ideal. > > > > As I see reuiring aio in userland is for buffer pre-allocating and > > pining, please check me, this is already true for all in-kernel > > operations? > > Not quite. The NFS client just accepts whatever mbuf it gets with RPCs, it > doesn't preallocate buffers specifically that are copied into (OTOH, this > means that when NFS handles things metadata RPCs it is already "zero-copy"). > Also, because of the framing and the fact that you can't control the order in > which RPCs are replied to, you can't queue buffers belonging to a file direct > for DMA from the NIC (for example). Oh, nfs/iscsi have additional framing, over TCP. This is cant't be simple handling. Intel/chelsio have NFS/iSCSI offload, but this is more complex, yes?
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20160509193137.GH1447>