Date: Wed, 1 Feb 2023 14:33:32 -0500 From: Paul Mather <paul@gromit.dlib.vt.edu> To: Marek Zarychta <zarychtam@plan-b.pwste.edu.pl> Cc: stable@freebsd.org Subject: Re: Slow WAN traffic to FreeBSD hosts but not to Linux hosts---how to debug/fix? Message-ID: <83E43236-60F8-4949-8840-54E66D327EE9@gromit.dlib.vt.edu> In-Reply-To: <8AE3B49C-6C7F-4A20-B2DC-0D4B1343FB59@gromit.dlib.vt.edu> References: <95EDCFCA-7E3F-458F-85A6-856D606B9D98@gromit.dlib.vt.edu> <4ed8b724-041f-f561-ae60-ab966aefbb68@plan-b.pwste.edu.pl> <282AF730-E5E0-4A50-9F47-E7301B36E5C8@gromit.dlib.vt.edu> <2ed582b9-b544-74bb-2047-99d04924b46b@plan-b.pwste.edu.pl> <8AE3B49C-6C7F-4A20-B2DC-0D4B1343FB59@gromit.dlib.vt.edu>
next in thread | previous in thread | raw e-mail | index | archive | help
On Feb 1, 2023, at 11:26 AM, Paul Mather <paul@gromit.dlib.vt.edu> = wrote: > On Jan 31, 2023, at 3:38 PM, Marek Zarychta = <zarychtam@plan-b.pwste.edu.pl> wrote: >=20 >> W dniu 31.01.2023 o 19:31, Paul Mather pisze: >>>> While playing with different mod_cc(4) might bring some = improvement, to get a real boost I'd suggest enabling tcp_rack(4) if = feasible. >>>=20 >>> I am interested in trying this out, but believe it is more feasible = in my case for the -STABLE and -CURRENT systems I am using, not so much = for the -RELEASE systems that are kept up to date via binary = freebsd-update updates. My reading of the tcp_rack(4) man page is that = you have to build a custom kernel as, unlike the cc_* congestion control = algorithms, the loadable tcp_rack module is not built by default. Is = that an accurate reading? >>>=20 >> Yes, this gift from Netflix is probably better suited for -STABLE and = -CURRENT as easier to set up there. There is an excellent, up-to-date = article about it by Klara Systems writers[1]. =46rom my experience = tcp_rack(4) is well suited for congested, lossy or redundant network = paths where loses, duplicated packets or races between packets occur. = Not a panacea, but very performant TCP stack based on the _fair_ = algorithm. In some instances, it might help you to saturate the = bandwidth of the link. TCP algo can be loaded/unloaded/changed on the = fly. In FreeBSD 14-CURRENT you can change it on an active socket with = tcpsso(8) utility, in FreeBSD 12 and 13 you have to restart the app = bound to the socket. >> Please feel free to play with TCP stacks and congestion algos with = the help of benchmarks/iperf3 to find out what prevents the link from = being saturated and give us some feedback here. >>=20 >> [1] = https://klarasystems.com/articles/using-the-freebsd-rack-tcp-stack/ >=20 >=20 > Thank you, Marek, for the link to the Klara article about building and = using RACK. I'm building it now on a FreeBSD-CURRENT system and will = test it out. It looks like we may have a winner, folks. I built and enabled the = extra TCP stacks and for the first time was able to max out my = connection to the remote FreeBSD system. I get consistently higher = throughput over the 15-hop WAN path to the remote FreeBSD system when = using the RACK TCP stack than when using the default "freebsd" stack. Although the speeds are consistently higher when using the setting = "net.inet.tcp.functions_default=3Drack", they are still variable. = However, rather than the 3--4 MB/s I saw that kicked off this thread, I = now average over 10 MB/s. I actually get the best results with = "net.inet.tcp.functions_default=3Dbbr" (having loaded tcp_bbr). That = behaves very much like the Linux hosts in that speeds climb very quickly = until it saturates the WAN connection. I get the same high speeds from = the remote FreeBSD system using tcp_bbr as I do to the Linux hosts. I = will stick with tcp_bbr for now as the default on my remote FreeBSD = servers. It appears to put them on a par with Linux for this WAN link. Cheers, Paul.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?83E43236-60F8-4949-8840-54E66D327EE9>