[ofa-general] Re: [PATCH 00/12 -Rev2] Implement batching skb API

[Krishna Kumar2]

> I have started a 10 run test for various buffer sizes and processes, and
> will post the results on Monday.

The 10 iteration run results for Rev2 are (average) :

----------------------------------------------------------------------------------
Test Case               Org         New            %Change
----------------------------------------------------------------------------------

                          TCP 1 Process
Size:32               2703            3063            13.31
Size:128             12948          12217           -5.64
Size:512             48108          55384           15.12
Size:4096           129089        132586          2.70
Average:            192848         203250          5.39

                          TCP 4 Processes
Size:32              10389          10768            3.64
Size:128            39694           42265           6.47
Size:512            159563         156373         -1.99
Size:4096           268094         256008        -4.50
Average:            477740         465414         -2.58


                          TCP No Delay 1 Process
Size:32               2606           2950            13.20
Size:128             8115           11864           46.19
Size:512             39113          42608           8.93
Size:4096           103966        105333          1.31
Average:            153800         162755          5.82


                  TCP No Delay 4 Processes
Size:32               4213            8727            107.14
Size:128             17579           35143           99.91
Size:512             70803           123936         75.04
Size:4096           203541          225259         10.67
Average:             296136         393065          32.73

--------------------------------------------------------------------------
Average:            1120524        1224484         9.28%

There are three cases that degrade a little (upto -5.6%), but there are 13
cases
that improve, and many of those are in the 13% to over 100% (7 cases).

Thanks,

- KK

Krishna Kumar2/India/IBM at IBMIN wrote on 07/22/2007 02:34:57 PM:

> This set of patches implements the batching API, and makes the following
> changes resulting from the review of the first set:
>
> Changes :
> ---------
> 1.  Changed skb_blist from pointer to static as it saves only 12 bytes
>     (i386), but bloats the code.
> 2.  Removed requirement for driver to set "features & NETIF_F_BATCH_SKBS"
>     in register_netdev to enable batching as it is redundant. Changed
this
>     flag to NETIF_F_BATCH_ON and it is set by register_netdev, and other
>     user changable calls can modify this bit to enable/disable batching.
> 3.  Added ethtool support to enable/disable batching (not tested).
> 4.  Added rtnetlink support to enable/disable batching (not tested).
> 5.  Removed MIN_QUEUE_LEN_BATCH for batching as high performance drivers
>     should not have a small queue anyway (adding bloat).
> 6.  skbs are purged from dev_deactivate instead of from unregister_netdev
>     to drop all references to the device.
> 7.  Removed changelog in source code in sch_generic.c, and unrelated
renames
>     from sch_generic.c (lockless, comments).
> 8.  Removed xmit_slots entirely, as it was adding bloat (code and header)
>     and not adding value (it is calculated and set twice in internal send
>     routine and handle work completion, and referenced once in batch
xmit;
>     and can instead be calculated once in xmit).
>
> Issues :
> --------
> 1. Remove /sysfs support completely ?
> 2. Whether rtnetlink support is required as GSO has only ethtool ?
>
> Patches are described as:
>    Mail 0/12  : This mail.
>    Mail 1/12  : HOWTO documentation.
>    Mail 2/12  : Changes to netdevice.h
>    Mail 3/12  : dev.c changes.
>    Mail 4/12  : Ethtool changes.
>    Mail 5/12  : sysfs changes.
>    Mail 6/12  : rtnetlink changes.
>    Mail 7/12  : Change in qdisc_run & qdisc_restart API, modify callers
>            to use this API.
>    Mail 8/12  : IPoIB include file changes.
>    Mail 9/12  : IPoIB verbs changes
>    Mail 10/12 : IPoIB multicast, CM changes
>    Mail 11/12 : IPoIB xmit API addition
>    Mail 12/12 : IPoIB xmit internals changes (ipoib_ib.c)
>
> I have started a 10 run test for various buffer sizes and processes, and
> will post the results on Monday.
>
> Please review and provide feedback/ideas; and consider for inclusion.
>
> Thanks,
>
> - KK