they told me that slow VirtIO drivers for FreeBSD are a known issue. Not a big deal then, KVM was developed on Linux and sure Linux guest drivers are more optimized.
bhyve also provides virtio devices, so this is important even without any Linuxes involved.
11.2-RELEASE-p4
I’d like to see 12.0-BETA4 in the comparison.. though most performance improvements might’ve been merged into 11..
as written in the FreeBSD network optimization guide
Not the guide you should be looking at. calomel.org > wiki.
No, stop. NEVER give calomel.org any traffic for any reason whatsoever. Do not send people there. Do not recommend anything on that site. Stop. Please.
If you even mention Calomel on the OpenBSD mailing lists, for example, the result is usually blacklisting or banning. The information on that site is usually quite inaccurate and/or out of date. Bad things happen when people follow guides on there especially as I’ve seen tuning suggestions listed that the author has no idea what they do.
Just avoid it. Please. If you want accurate information come to the mailing lists after our documentation has failed to provide you with the info you need.
Very little tuning information supplied (a few sysctl values, mention of turning off meltdown/spectre). Were default installs compared (aside from the tidbits mentioned)? Or both systems further tuned somehow?
As an example, FreeBSD can generally benefit greatly (see refs) from tuning vnodes and nic/network tuning (hw rxd/txd/rx_process_limit/ifqmaxlen values, disabling network based entropy collection). For what it is worth, the author is apparently a RedHat employee and a Linux kernel contributor, so I would assume they have a decent knowledge of tuning on the Linux side – no mention if any additional tuning was done though, so perhaps not.
Cross platform comparison benchmarking is really hard to do thoroughly and well.
So results are interesting, insofar as network benchmarks/comparisons are generally interesting.
I am a kernel contributor, but the Linux system was not tuned at all, the only “tuning” on Linux was to unload the iptables module in the routing benchmark, as Linux doesn’t have a sysctl to do it.
For the FreeBSD system, it was fully stock too, just with entropy harvesting disabled.
bhyve also provides virtio devices, so this is important even without any Linuxes involved.
I’d like to see 12.0-BETA4 in the comparison.. though most performance improvements might’ve been merged into 11..
Not the guide you should be looking at. calomel.org > wiki.
I think the fast-forwarding changes (if that is what you were thinking of) got merged into 11.x already.
From https://wiki.freebsd.org/NetworkPerformanceTuning :
No, stop. NEVER give calomel.org any traffic for any reason whatsoever. Do not send people there. Do not recommend anything on that site. Stop. Please.
Uh, okay?? Would be nice to hear actual reasons though, not just “no stop never”. The guide was pretty helpful for me.
Articles on the bsd router project are generally pretty high quality and usually have a description of what the tunable does.
I will have a look
If you even mention Calomel on the OpenBSD mailing lists, for example, the result is usually blacklisting or banning. The information on that site is usually quite inaccurate and/or out of date. Bad things happen when people follow guides on there especially as I’ve seen tuning suggestions listed that the author has no idea what they do.
Just avoid it. Please. If you want accurate information come to the mailing lists after our documentation has failed to provide you with the info you need.
Very little tuning information supplied (a few sysctl values, mention of turning off meltdown/spectre). Were default installs compared (aside from the tidbits mentioned)? Or both systems further tuned somehow?
As an example, FreeBSD can generally benefit greatly (see refs) from tuning vnodes and nic/network tuning (hw rxd/txd/rx_process_limit/ifqmaxlen values, disabling network based entropy collection). For what it is worth, the author is apparently a RedHat employee and a Linux kernel contributor, so I would assume they have a decent knowledge of tuning on the Linux side – no mention if any additional tuning was done though, so perhaps not.
Cross platform comparison benchmarking is really hard to do thoroughly and well.
So results are interesting, insofar as network benchmarks/comparisons are generally interesting.
refs:
Hi all,
I am a kernel contributor, but the Linux system was not tuned at all, the only “tuning” on Linux was to unload the iptables module in the routing benchmark, as Linux doesn’t have a sysctl to do it. For the FreeBSD system, it was fully stock too, just with entropy harvesting disabled.
Entropy collection was actually mentioned. But not things like mbufs.
netstat -noutput would be nice.Ah. I must have missed the part about entropy collection.