Date: Mon, 13 Sep 2010 14:18:28 -0500 From: Tom Judge <tom@tomjudge.com> To: Andre Oppermann <andre@freebsd.org> Cc: pyunyh@gmail.com, freebsd-net@freebsd.org, davidch@broadcom.com, yongari@freebsd.org Subject: Re: bce(4) - com_no_buffers (Again) Message-ID: <4C8E7904.9090004@tomjudge.com> In-Reply-To: <4C8E775D.8070202@freebsd.org> References: <4C894A76.5040200@tomjudge.com> <20100910002439.GO7203@michelle.cdnetworks.com> <4C8E3D79.6090102@tomjudge.com> <20100913184833.GF1229@michelle.cdnetworks.com> <4C8E775D.8070202@freebsd.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On 09/13/2010 02:11 PM, Andre Oppermann wrote: > On 13.09.2010 20:48, Pyun YongHyeon wrote: >> On Mon, Sep 13, 2010 at 10:04:25AM -0500, Tom Judge wrote: >>> Without BCE_JUMBO_HDRSPLIT then we see no errors. With it we see >>> number >>> of errors, however the rate seems to be reduced compaired to the >>> previous version of the driver. Please note that 'rate' here relates to the rate at which dev.bce.X.com_no_buffers is increasing not to PPS or bandwidth. However the discussion is still interesting. >>> >> >> It seems there are issues in header splitting and it was disabled >> by default. Header splitting reduces packet processing overhead in >> upper layer so it's normal to see better performance with header >> splitting. > > I'm not sure that header splitting really helps much at least for TCP. > The only place where it could make a difference is at socket buffer > append time. There the header get 'thrown away'. With header splitting > the first mbuf in the chain containing the header can be returned to the > free pool. Without header splitting it's just a offset change in the > mbuf. > > IIRC header splitting was introduced with the Tigeon cards which were > the first programmable network cards and the first to support putting > the header in a different mbuf. Header splitting, in theory, could > make a difference with zero copy sockets where the data portion in a > separate mbuf is flipped by VM magic into userspace. The trouble is > that no driver fully supports the semantics required for page flipping > and the zero copy code, if compiled in, is less much less optimized for > the non-flipping case than the standard code path. With the many dozen > gigabyte per second memory copy bandwidth of current CPU's it remains > questionable whether the page-flipping VM magic is actually faster than > a plain kernel/userspace copy as in the standard code path. I generally > recommend not to use ZERO_COPY_SOCKETS. > > I suspect in the case of the bce(4) driver the change in header splitting > is probably not the cause of the performance difference. > -- TJU13-ARIN
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?4C8E7904.9090004>