Skip site navigation (1)Skip section navigation (2)
Date:      Thu, 16 Feb 2017 23:24:57 +0100
From:      Sydney Meyer <meyer.sydney@googlemail.com>
To:        freebsd-net@freebsd.org
Subject:   Re: Slow Download Speeds from AWS S3
Message-ID:  <5BE01E68-A8EF-4572-9B48-C0C58683F085@googlemail.com>
In-Reply-To: <673aca75-7bb5-73e7-d23b-b12a25816c8a@multiplay.co.uk>
References:  <A6482775-D7F8-4FB8-9423-257D6D625D01@googlemail.com> <18e49da9-47eb-1803-e223-b4385e7d8690@multiplay.co.uk> <30F9FC69-CEBE-4CBE-93A9-DEDBD053C6C6@googlemail.com> <673aca75-7bb5-73e7-d23b-b12a25816c8a@multiplay.co.uk>

next in thread | previous in thread | raw e-mail | index | archive | help
Hi Steve,

> On 16 Feb 2017, at 18:18, Steven Hartland <killing@multiplay.co.uk> =
wrote:
>=20
> It does seem to be related to TCP Receive Window Size.
>=20
> When I tested on a 11.0-RELEASE box I got 30MB/s out of the box and =
the increase of recvspace was only needed on the original 10.2 however =
the 11 box is only 1.2ms from AWS where as the 10.2 was 17ms away, so =
likely explains the difference.

Yes, i saw this too.. when i did some tests with a VM from within the =
same AWS region, i did not see the troughput issues as with connections =
with a higher latency (probably should have mentioned this in the first =
post, retrospectively).. this would also explain the vmware guest =
bridged vs nat'ed difference then..

>=20
> It seems that when talking to hosts that don't support Timestamps =
FreeBSD's performance is going to be poor due to lack of auto scaling :(

I did some tests with Debian 8.x (Linux Kernel 3.16.x) and there the =
default for /proc/sys/net/ipv4/tcp_timestamps appears to be enabled, at =
least with Debian's 3.16 Linux Kernel. I can see the TS val field now in =
tcpdump which is lacking in traffic from AWS S3, regardless from region.

>=20
> Reading the thread you linked is seems like we need to implement auto =
scaling based on RTT estimations similar to linux.

Well, i guess the issue described in the related thread is indeed the =
case here.
I will adjust this buffer size manually on this machine, as it has no =
other major purpose than pulling archives from S3 anyway.

But thanks again, learned something new (to me) about FreeBSDs =
networking stack.

Updated the Forums Thread accordingly here:

https://forums.freebsd.org/threads/59756/

Cheers,

Sydney.

>=20
>    Regards
>    Steve
>=20
> On 16/02/2017 16:55, Sydney Meyer via freebsd-net wrote:
>> Hi Steve,
>>=20
>> increasing the buffer size did indeed enhance throughput.
>>=20
>> I am obviously not an expert in this field, but i don't understand =
why or if the TCP Receive Window Size shouldn't increase automatically.
>>=20
>> I found this thread on the ML and i'm reading myself a bit more into =
the topic right now.
>>=20
>> =
https://lists.freebsd.org/pipermail/freebsd-net/2014-August/039495.html
>>=20
>> Thank you for your tip, however..
>>=20
>> Sydney
>>=20
>>> On 16 Feb 2017, at 16:35, Steven Hartland <killing@multiplay.co.uk> =
wrote:
>>>=20
>>> Window scaling and receive buffer scaling is the most likely cause.
>>>=20
>>> Check what the sysctl net.inet.tcp.recvspace is set to, then try =
increasing it e.g.
>>> sysctl net.inet.tcp.recvspace=3D655360
>>>=20
>>> This jumped the transfer rate with a wget and your test URL from =
3MB/s to 30MB/s here.
>>>=20
>>>    Regards
>>>    Steve
>>>=20
>>> On 16/02/2017 14:34, Sydney Meyer via freebsd-net wrote:
>>>> Hello,
>>>>=20
>>>> I'm seeing continuous slow download speeds from Amazon S3, but only =
on FreeBSD. Other OSes saturate the connection without problems.
>>>>=20
>>>> This happens with 10.3-RELEASE and 11.0-RELEASE and only with AWS =
S3 in different regions (Ireland, London, Frankfurt, US Standard have =
been tested) whilst using fetch, curl, et. al.
>>>>=20
>>>> I have tested this on multiple machines, bare metal, bhyve, Xen and =
VMware VMs, routed setups and direct pppoe links.. all the same.
>>>>=20
>>>> Anyone seeing similar issues?
>>>>=20
>>>> Here's a url to try:
>>>>=20
>>>> =
http://s3.eu-central-1.amazonaws.com/4f48caf1d8bcbef8/c5b38f8b3625d2b6/zer=
ofile.raw
>>>>=20
>>>> Also, when doing double NAT, i.e. VMware Fusion FreeBSD Guest with =
"Share with my Mac"-Interface, the machine is doing completely fine, as =
in saturates the link, 90Mbps otherwise between 12-15Mbps..
>>>>=20
>>>> I also switched the FreeBSD Routers with Linux-based ones and with =
the isp-provided router box, with the same result.
>>>>=20
>>>> I have launched VMs with Digital Ocean to "rule out" my ISP and =
there seems to be the same issue. Downloading from S3 is multiple times =
slower than any other connection to services outside of S3 or on any =
other OS.
>>>>=20
>>>> It seems like other people are seeing the same issue:
>>>>=20
>>>> https://forums.freebsd.org/threads/59756/#post-343064
>>>> _______________________________________________
>>>> freebsd-net@freebsd.org mailing list
>>>> https://lists.freebsd.org/mailman/listinfo/freebsd-net
>>>> To unsubscribe, send any mail to =
"freebsd-net-unsubscribe@freebsd.org"
>>> _______________________________________________
>>> freebsd-net@freebsd.org mailing list
>>> https://lists.freebsd.org/mailman/listinfo/freebsd-net
>>> To unsubscribe, send any mail to =
"freebsd-net-unsubscribe@freebsd.org"
>> _______________________________________________
>> freebsd-net@freebsd.org mailing list
>> https://lists.freebsd.org/mailman/listinfo/freebsd-net
>> To unsubscribe, send any mail to =
"freebsd-net-unsubscribe@freebsd.org"
>=20
> _______________________________________________
> freebsd-net@freebsd.org mailing list
> https://lists.freebsd.org/mailman/listinfo/freebsd-net
> To unsubscribe, send any mail to "freebsd-net-unsubscribe@freebsd.org"




Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?5BE01E68-A8EF-4572-9B48-C0C58683F085>