[Cerowrt-devel] Fwd: Poor TCP performance with ath10k in 4.0 kernel, again.

[Ben Greear]

On 05/19/2015 01:49 PM, Dave Taht wrote:
> ben, do you have packet captures?

I can get them easily enough...what scenario would you like
captured?  Files will be large, but I can upload them somewhere
and post a link.

> 
> What was the qdisc on the interface?

Default kernel qdisc...I think fifo-fast?

Thanks,
Ben

> 
> ---------- Forwarded message ----------
> From: Ben Greear <greearb at candelatech.com>
> Date: Mon, May 18, 2015 at 2:27 PM
> Subject: Poor TCP performance with ath10k in 4.0 kernel, again.
> To: "linux-wireless at vger.kernel.org" <linux-wireless at vger.kernel.org>,
> ath10k <ath10k at lists.infradead.org>
> 
> 
> Disclosure: I am working with a patched 4.0 kernel, patched ath10k driver, and
> patched (CT) ath10k firmware.  Traffic generator is of our own making.
> 
> First, this general problem has been reported before, but the
> work-arounds previously suggested do not fully resolve my problems.
> 
> The basic issue is that when the sending socket is directly on top
> of a wifi interface (ath10k driver), then TCP throughput sucks.
> 
> For instance, if AP interface sends to station, with 10 concurrent
> TCP streams, I see about 426Mbps.  With 100 streams, I see total throughput
> of 750Mbps.  These were maybe 10-30 second tests that I did.
> 
> Interestingly, a single stream connection performs very poorly at first,
> but at least in one test, it eventually ran quite fast.  It is too
> complicated to describe in words, but the graph is here:
> 
> http://www.candelatech.com/downloads/single-tcp-4.0.pdf
> 
> The 10-stream test did not go above about 450Mbps even after running
> for more than
> 1 minute, and it was fairly stable around the 450Mbps range after the
> first few seconds.
> 
> 100-stream test shows nice stable aggregate throughput:
> 
> http://www.candelatech.com/downloads/100-tcp-4.0.pdf
> 
> I have tweaked the kernel tcp_limit_output_bytes setting
> (tested at 1024k too, did not make any significant difference).
> 
> # cat /proc/sys/net/ipv4/tcp_limit_output_bytes
> 2048000
> 
> I have tried forcing TCP send/rcv buffers to be 1MB and 2MB, but that
> did not make obvious difference except that it started at the maximum
> rate very quickly instead of taking a few seconds to train up to full speed.
> 
> If I run a single-stream TCP test, sending on eth1 (Intel 1G NIC)
> through the AP machine, then single stream download is about 540 Mbps,
> and ramps up
> quickly.  So, the AP can definitely send the needed amount of TCP packets.
> 
> UDP throughput in download direction, single stream, is about 770Mbps,
> regardless
> of whether I originate the socket on the AP or if I pass it through the AP.
> send/recv bufs are set to 1MB for UDP sockets.
> 
> The 3.17 kernel shows similar behaviour, and the 3.14 kernel is a lot better
> for TCP traffic.
> 
> Are there tweaks other than tcp_limit_output_bytes that might
> improve this behaviour?
> 
> I will be happy to grab captures or provide any other debugging info
> that someone thinks will be helpful.
> 
> Thanks,
> Ben
> 
> --
> Ben Greear <greearb at candelatech.com>
> Candela Technologies Inc  http://www.candelatech.com
> 
> 
> _______________________________________________
> ath10k mailing list
> ath10k at lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/ath10k
> 
> 


-- 
Ben Greear <greearb at candelatech.com>
Candela Technologies Inc  http://www.candelatech.com