[Bloat] Excessive throttling with fq
Hans-Kristian Bakke
hkbakke at gmail.com
Thu Jan 26 15:57:42 EST 2017
The receiver (as in the nginx proxy in the dumps) is actually running fq
qdisc with BBR on kernel 4.9. Could that explain what you are seeing?
Changing it to cubic does not change the resulting throughput though, and
it also was not involved at all in the Windows 10 -> linux router -> Apache
server tests which also gives the same 23-ish MB/s with pacing.
On 26 January 2017 at 21:54, Eric Dumazet <eric.dumazet at gmail.com> wrote:
> It looks like the receiver is seriously limiting receive window in the
> "pacing" case.
>
> Only after 30 Mbytes are transfered, it finally increases it, from
> 359168 to 1328192
>
> DRS is not working as expected. Again maybe related to HZ value.
>
>
> On Thu, 2017-01-26 at 21:19 +0100, Hans-Kristian Bakke wrote:
> > There are two packet captures from fq with and without pacing here:
> >
> >
> > https://owncloud.proikt.com/index.php/s/KuXIl8h8bSFH1fM
> >
> >
> >
> > The server (with fq pacing/nopacing) is 10.0.5.10 and is running a
> > Apache2 webserver at port tcp port 443. The tcp client is nginx
> > reverse proxy at 10.0.5.13 on the same subnet which again is proxying
> > the connection from the Windows 10 client.
> > - I did try to connect directly to the server with the client (via a
> > linux gateway router) avoiding the nginx proxy and just using plain
> > no-ssl http. That did not change anything.
> > - I also tried stopping the eth0 interface to force the traffic to the
> > eth1 interface in the LACP which changed nothing.
> > - I also pulled each of the cable on the switch to force the traffic
> > to switch between interfaces in the LACP link between the client
> > switch and the server switch.
> >
> >
> > The CPU is a 5-6 year old Intel Xeon X3430 CPU @ 4x2.40GHz on a
> > SuperMicro platform. It is not very loaded and the results are always
> > in the same ballpark with fq pacing on.
> >
> >
> >
> > top - 21:12:38 up 12 days, 11:08, 4 users, load average: 0.56, 0.68,
> > 0.77
> > Tasks: 1344 total, 1 running, 1343 sleeping, 0 stopped, 0 zombie
> > %Cpu0 : 0.0 us, 1.0 sy, 0.0 ni, 99.0 id, 0.0 wa, 0.0 hi, 0.0
> > si, 0.0 st
> > %Cpu1 : 0.0 us, 0.3 sy, 0.0 ni, 97.4 id, 2.0 wa, 0.0 hi, 0.3
> > si, 0.0 st
> > %Cpu2 : 0.0 us, 2.0 sy, 0.0 ni, 96.4 id, 1.3 wa, 0.0 hi, 0.3
> > si, 0.0 st
> > %Cpu3 : 0.7 us, 2.3 sy, 0.0 ni, 94.1 id, 3.0 wa, 0.0 hi, 0.0
> > si, 0.0 st
> > KiB Mem : 16427572 total, 173712 free, 9739976 used, 6513884
> > buff/cache
> > KiB Swap: 6369276 total, 6126736 free, 242540 used. 6224836 avail
> > Mem
> >
> >
> > This seems OK to me. It does have 24 drives in 3 ZFS pools at 144TB
> > raw storage in total with several SAS HBAs that is pretty much always
> > poking the system in some way or the other.
> >
> >
> > There are around 32K interrupts when running @23 MB/s (as seen in
> > chrome downloads) with pacing on and about 25K interrupts when running
> > @105 MB/s with fq nopacing. Is that normal?
> >
> >
> > Hans-Kristian
> >
> >
> >
> > On 26 January 2017 at 20:58, David Lang <david at lang.hm> wrote:
> > Is there any CPU bottleneck?
> >
> > pacing causing this sort of problem makes me thing that the
> > CPU either can't keep up or that something (Hz setting type of
> > thing) is delaying when the CPU can get used.
> >
> > It's not clear from the posts if the problem is with sending
> > data or receiving data.
> >
> > David Lang
> >
> >
> > On Thu, 26 Jan 2017, Eric Dumazet wrote:
> >
> > Nothing jumps on my head.
> >
> > We use FQ on links varying from 1Gbit to 100Gbit, and
> > we have no such
> > issues.
> >
> > You could probably check on the server the TCP various
> > infos given by ss
> > command
> >
> >
> > ss -temoi dst <remoteip>
> >
> >
> > pacing rate is shown. You might have some issues, but
> > it is hard to say.
> >
> >
> > On Thu, 2017-01-26 at 19:55 +0100, Hans-Kristian Bakke
> > wrote:
> > After some more testing I see that if I
> > disable fq pacing the
> > performance is restored to the expected
> > levels: # for i in eth0 eth1; do tc qdisc
> > replace dev $i root fq nopacing;
> > done
> >
> >
> > Is this expected behaviour? There is some
> > background traffic, but only
> > in the sub 100 mbit/s on the switches and
> > gateway between the server
> > and client.
> >
> >
> > The chain:
> > Windows 10 client -> 1000 mbit/s -> switch ->
> > 2xgigabit LACP -> switch
> > -> 4 x gigabit LACP -> gw (fq_codel on all
> > nics) -> 4 x gigabit LACP
> > (the same as in) -> switch -> 2 x lacp ->
> > server (with misbehaving fq
> > pacing)
> >
> >
> >
> > On 26 January 2017 at 19:38, Hans-Kristian
> > Bakke <hkbakke at gmail.com>
> > wrote:
> > I can add that this is without BBR,
> > just plain old kernel 4.8
> > cubic.
> >
> > On 26 January 2017 at 19:36,
> > Hans-Kristian Bakke
> > <hkbakke at gmail.com> wrote:
> > Another day, another fq issue
> > (or user error).
> >
> >
> > I try to do the seeminlig
> > simple task of downloading a
> > single large file over local
> > gigabit LAN from a
> > physical server running kernel
> > 4.8 and sch_fq on intel
> > server NICs.
> >
> >
> > For some reason it wouldn't go
> > past around 25 MB/s.
> > After having replaced SSL with
> > no SSL, replaced apache
> > with nginx and verified that
> > there is plenty of
> > bandwith available between my
> > client and the server I
> > tried to change qdisc from fq
> > to pfifo_fast. It
> > instantly shot up to around
> > the expected 85-90 MB/s.
> > The same happened with
> > fq_codel in place of fq.
> >
> >
> > I then checked the statistics
> > for fq and the throttled
> > counter is increasing
> > massively every second (eth0 and
> > eth1 is LACPed using Linux
> > bonding so both is seen
> > here):
> >
> >
> > qdisc fq 8007: root refcnt 2
> > limit 10000p flow_limit
> > 100p buckets 1024 orphan_mask
> > 1023 quantum 3028
> > initial_quantum 15140
> > refill_delay 40.0ms
> > Sent 787131797 bytes 520082
> > pkt (dropped 15,
> > overlimits 0 requeues 0)
> > backlog 98410b 65p requeues 0
> > 15 flows (14 inactive, 1
> > throttled)
> > 0 gc, 2 highprio, 259920
> > throttled, 15 flows_plimit
> > qdisc fq 8008: root refcnt 2
> > limit 10000p flow_limit
> > 100p buckets 1024 orphan_mask
> > 1023 quantum 3028
> > initial_quantum 15140
> > refill_delay 40.0ms
> > Sent 2533167 bytes 6731 pkt
> > (dropped 0, overlimits 0
> > requeues 0)
> > backlog 0b 0p requeues 0
> > 24 flows (24 inactive, 0
> > throttled)
> > 0 gc, 2 highprio, 397
> > throttled
> >
> >
> > Do you have any suggestions?
> >
> >
> > Regards,
> > Hans-Kristian
> >
> >
> >
> >
> > _______________________________________________
> > Bloat mailing list
> > Bloat at lists.bufferbloat.net
> > https://lists.bufferbloat.net/listinfo/bloat
> >
> >
> > _______________________________________________
> > Bloat mailing list
> > Bloat at lists.bufferbloat.net
> > https://lists.bufferbloat.net/listinfo/bloat
> >
> >
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.bufferbloat.net/pipermail/bloat/attachments/20170126/d4c91fa7/attachment-0002.html>
More information about the Bloat
mailing list