Re: tcp: add NV congestion control

Re: tcp: add NV congestion control

[Lawrence Brakmo]

The updated NV document with the new experiments and a table with all the
experimental results are now available at
(http://www.brakmo.org/networking/tcp-nv/TCPNV.html).

- Lawrence
 

On 8/25/15, 4:33 PM, "Lawrence Brakmo" <brakmo@fb.com> wrote:

>Changes from v5: cleaning of NV code, changing some default parameters
>
>I've run more extensive tests, I'm working on updating the NV website
>(http://www.brakmo.org/networking/tcp-nv/TCPNV.html) should be updated
>by tomorrow (8/26).
>
>The updated tests include Reno, Cubic, NV and CDG and include more types
>of traffic. Overview of results:
>1) NV has a little lower throughput (2-3% less) with small number of flows
>   as compared to Reno, Cubic and CDG
>2) NV is less fair with few flows but becomes more fair with more flows
>3) Less losses with NV (none in many cases) as compared to all others.
>   One exception is when things get very congested (64 flows into one
>   server), NV has 50% more losses than CDG, Cubic has 1.8x to 10x more
>   losses than CDG. Reno has about the same losses as CDG.
>4) In mixed traffic (1M and 10K RPCs), 10K flows achieve much higher
>   average throughput with NV than with the others (which are
>   very similar). In one example, 2 clients sending 1M and 10K to 2
>   servers, with NV 10K flows average 1Gbps and 1M flows 3.7Gbps,
>   whereas they average about 226Mbps and 4.4Gbps for Reno, Cubic and
>   CDG. They all have similar link utilization.
>
>Consists of the following patches:
>
>[RFC PATCH v6 net-next 1/4] tcp: replace cnt & rtt with struct in
>[RFC PATCH v6 net-next 2/4] tcp:  refactor struct tcp_skb_cb
>[RFC PATCH v6 net-next 3/4] tcp: add in_flight to tcp_skb_cb
>[RFC PATCH v6 net-next 4/4] tcp: add NV congestion control
>
>Signed-off-by: Lawrence Brakmo <brakmo@fb.com>
>
>include/net/tcp.h       |  20 ++-
>net/ipv4/Kconfig        |  16 ++
>net/ipv4/Makefile       |   1 +
>net/ipv4/tcp_bic.c      |   6 +-
>net/ipv4/tcp_cdg.c      |  14 +-
>net/ipv4/tcp_cubic.c    |   6 +-
>net/ipv4/tcp_htcp.c     |  10 +-
>net/ipv4/tcp_illinois.c |  20 +--
>net/ipv4/tcp_input.c    |  10 +-
>net/ipv4/tcp_lp.c       |   6 +-
>net/ipv4/tcp_nv.c       | 489
>++++++++++++++++++++++++++++++++++++++++++++++++++++++
>net/ipv4/tcp_output.c   |   4 +-
>net/ipv4/tcp_vegas.c    |   6 +-
>net/ipv4/tcp_vegas.h    |   2 +-
>net/ipv4/tcp_veno.c     |   7 +-
>net/ipv4/tcp_westwood.c |   7 +-
>net/ipv4/tcp_yeah.c     |   7 +-
>17 files changed, 580 insertions(+), 51 deletions(-)

tcp: add NV congestion control

[Lawrence Brakmo]

Changes from v5: cleaning of NV code, changing some default parameters

I've run more extensive tests, I'm working on updating the NV website
(http://www.brakmo.org/networking/tcp-nv/TCPNV.html) should be updated
by tomorrow (8/26).

The updated tests include Reno, Cubic, NV and CDG and include more types
of traffic. Overview of results:
1) NV has a little lower throughput (2-3% less) with small number of flows
   as compared to Reno, Cubic and CDG
2) NV is less fair with few flows but becomes more fair with more flows
3) Less losses with NV (none in many cases) as compared to all others.
   One exception is when things get very congested (64 flows into one
   server), NV has 50% more losses than CDG, Cubic has 1.8x to 10x more
   losses than CDG. Reno has about the same losses as CDG.
4) In mixed traffic (1M and 10K RPCs), 10K flows achieve much higher
   average throughput with NV than with the others (which are
   very similar). In one example, 2 clients sending 1M and 10K to 2
   servers, with NV 10K flows average 1Gbps and 1M flows 3.7Gbps,
   whereas they average about 226Mbps and 4.4Gbps for Reno, Cubic and
   CDG. They all have similar link utilization.

Consists of the following patches:

[RFC PATCH v6 net-next 1/4] tcp: replace cnt & rtt with struct in
[RFC PATCH v6 net-next 2/4] tcp:  refactor struct tcp_skb_cb
[RFC PATCH v6 net-next 3/4] tcp: add in_flight to tcp_skb_cb
[RFC PATCH v6 net-next 4/4] tcp: add NV congestion control

Signed-off-by: Lawrence Brakmo <brakmo@fb.com>

include/net/tcp.h       |  20 ++-
net/ipv4/Kconfig        |  16 ++
net/ipv4/Makefile       |   1 +
net/ipv4/tcp_bic.c      |   6 +-
net/ipv4/tcp_cdg.c      |  14 +-
net/ipv4/tcp_cubic.c    |   6 +-
net/ipv4/tcp_htcp.c     |  10 +-
net/ipv4/tcp_illinois.c |  20 +--
net/ipv4/tcp_input.c    |  10 +-
net/ipv4/tcp_lp.c       |   6 +-
net/ipv4/tcp_nv.c       | 489 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
net/ipv4/tcp_output.c   |   4 +-
net/ipv4/tcp_vegas.c    |   6 +-
net/ipv4/tcp_vegas.h    |   2 +-
net/ipv4/tcp_veno.c     |   7 +-
net/ipv4/tcp_westwood.c |   7 +-
net/ipv4/tcp_yeah.c     |   7 +-
17 files changed, 580 insertions(+), 51 deletions(-)

Re: tcp: add NV congestion control

[David Miller]

From: Lawrence Brakmo <brakmo@fb.com>
Date: Tue, 25 Aug 2015 16:33:50 -0700

> Changes from v5: cleaning of NV code, changing some default parameters

I have no fundamental objections to this patch series.