[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Xen-devel] [PATCH 00/22] Netchannel2 patches for XCI 2.6.27

This patch series contains all of the Linux-side changes from the
netchannel2 tree, ported to 2.6.27 plus the XCI patchqueue and tidied
up into a bunch of reasonably self-contained patches.

Headline features:

-- A new grant tables implementation.  This is both faster and more
   flexible, adding support for sub-page and transitive grants.
-- Receiver-side copying of packets, which gives better performance
   and better CPU accounting.
-- Domain 0 bypass support.  Domains which are located on the same
   physical host can set up a direct communication channel between
   themselves, improving performance and dramatically reducing the
   load on domain 0.  Bypasses are created and destroyed automatically
   in response to observed traffic patterns, and do not require manual
-- VMQ support, allowing guests to take advantage of virtual-machine
   acceleration support present in some modern NICs (but note the
   caveats below).

Performance is generally decent, in some cases achieving more than
twice the throughput of netchannel1, but this is rather
workload-dependent.  It shouldn't be significantly slower in any
realistic tests.

The patch series contains four main chunks:

-- First, a mechanism for netback to register its mappings of foreign
   pages in a way which will be accessible to netback2.  This is
   necessary when forwarding packets between the two backends.
   (patches 1 and 2 in the series)

-- Second, V2 grant tables.
   (patches 3 through 8)

-- Third, netchannel2 itself.
   (patches 9 through 18)

-- Finally, VMQ support for netchannel2.  The actual netchannel2 parts
   of this are reasonably okay, but in order to make use of it you
   need to have a VMQ-capable NIC with a VMQ-capable driver.  The
   final patch adds this capability to the ixgbe driver, but it's to a
   slightly out-of-date version of it.  Unfortunately, I don't
   currently have access to any VMQ-capable hardware, and the patch is
   too big to try to forward-port blind, so there's not a great deal I
   can do about this.
   (patches 19 through 22)

The series should apply cleanly to commit
59270a8c95901eaa9e3f06d575d4168c320ae057 of the XenClient linux-2.6.27
tree plus 925d65d3d407a620dffb86fb43bb84d9eec3cb07 of their

As far as I'm concerned, the first three chunks (everything up to
patch 18, ``Add support for automatically creating and destroying
bypass rings...'') is ready to go in to the XCI patchqueue.


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.