On Wed, 2017-09-13 at 19:27 +0300, Denys Fedoryshchenko wrote: > On 2017-09-13 19:16, Eric Dumazet wrote: > > On Wed, 2017-09-13 at 18:34 +0300, Denys Fedoryshchenko wrote: > >> Well, probably i am answering my own question, removing estimator from > >> classes seems drastically improve situation. > >> It seems estimator has some issues that cause shaper to behave > >> incorrectly (throttling traffic while it should not). > >> But i guess thats a bug? > >> As i was not able to predict such bottleneck by CPU load measurements. > > > > Well, there was a reason we disabled HTB class estimators by default ;) > > > > > > https://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next.git/commit/?id=64153ce0a7b61b2a5cacb01805cbf670142339e9 > > As soon as disabling it solve my problem - i'm fine, hehe, but i guess > other people who might hit this problem, should be aware how to find > reason. > They should not be disappointed in Linux :)
Well, if they enable rate estimators while kernel does not set them by default, they get what they want, at a cost. > Because i can't measure this bottleneck before it happens, i'm seeing on > mpstat all cpu's are idle, and same time traffic is throttled. Normally things were supposed to get much better in linux-4.10 ( https://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next.git/commit/?id=1c0d32fde5bdf1184bc274f864c09799278a1114 ) But I apparently added a scaling bug. I will try : diff --git a/net/core/gen_estimator.c b/net/core/gen_estimator.c index 0385dece1f6fe5e26df1ce5f40956a79a2eebbf4..7c1ffd6f950172c1915d8e5fa2b5e3f77e4f4c78 100644 --- a/net/core/gen_estimator.c +++ b/net/core/gen_estimator.c @@ -83,10 +83,10 @@ static void est_timer(unsigned long arg) u64 rate, brate; est_fetch_counters(est, &b); - brate = (b.bytes - est->last_bytes) << (8 - est->ewma_log); + brate = (b.bytes - est->last_bytes) << (10 - est->ewma_log - est->intvl_log); brate -= (est->avbps >> est->ewma_log); - rate = (u64)(b.packets - est->last_packets) << (8 - est->ewma_log); + rate = (u64)(b.packets - est->last_packets) << (10 - est->ewma_log - est->intvl_log); rate -= (est->avpps >> est->ewma_log); write_seqcount_begin(&est->seq); diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c