Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 8 Sep 2017 01:04:00 +1000
From:      Paul Koch <paul.koch@akips.com>
To:        Sepherosa Ziehau <sephe@freebsd.org>
Cc:        Julian Elischer <julian@freebsd.org>, freebsd-stable@freebsd.org, Hongjiang Zhang <honzhan@microsoft.com>
Subject:   Re: 11.1 running on HyperV hn interface hangs
Message-ID:  <20170908010400.0fb81bc4@akips.com>
In-Reply-To: <CAMOc5cx-dES8ABM8oNNKfT50ZUebVtpkKJn9Vtc_nQrRQ2vHgQ@mail.gmail.com>
References:  <20170906193309.796c79ed@akips.com> <3f96c7d0-4fbd-26cb-5c84-8868d12eb427@ingresso.co.uk> <14997bba-aac2-947b-9b78-04af41ea29b0@freebsd.org> <CAMOc5cxQp1P3WnDf=Hgo8X0LmQfjCTowO=ya8VsWhu4cdTeajQ@mail.gmail.com> <20170907150711.025e4e41@akips.com> <CAMOc5cx-dES8ABM8oNNKfT50ZUebVtpkKJn9Vtc_nQrRQ2vHgQ@mail.gmail.com>

next in thread | previous in thread | raw e-mail | index | archive | help
On Thu, 7 Sep 2017 13:51:11 +0800
Sepherosa Ziehau <sephe@freebsd.org> wrote:

> Weird, your traffic pattern does not even belong to anything heavy.
> Sending is mainly UDP, which will never be able to saturate the TX
> buffer ring causing the RXBUF ACK sending failure.  This is weird.

It's a bit tricky. The poller is very fast. We ping every device every 15
seconds, and collect every MIB object every 60 seconds. The poller "rate
limits" itself by dividing each minute into 100ms time slots and only sends a
specific amount of pings/snmp packets in each time slot.  The problem is, it
blasts the request packets out really fast at the start of each time slot,
and then sits in a receive loop until the next time slot comes around.  The
requests are not paced over the 100ms, therefore it will blast out a lot
of packets in a few milliseconds.

We use to use a 1 second rate limiting time slot, and didn't interlace
ping/snmp requests, but we found certain interface types on Cisco 6509
switches couldn't keep up with back-to-back pings and would lose them.


> Anyhow, make sure to test this patch:
> 8762017-Sep-07 02:19 hn_inc_txbr.diff

Yep.  Might take a bit of time to test though because we'll need to get the
customer to spin up a test VM on the same platform, and they are fairly
remote (Perth, Australia).  We don't run any Microsoft servers/HyperV setups
in our lab.

	Paul.
-- 
Paul Koch | Founder | CEO
AKIPS Network Monitor | akips.com
Brisbane, Australia



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20170908010400.0fb81bc4>