Date: Sat, 11 Dec 2021 14:45:31 -0500 From: Allan Jude <allanjude@freebsd.org> To: freebsd-current@freebsd.org Subject: Re: Benchmarks: FreeBSD 13 vs. NetBSD 9.2 vs. OpenBSD 7 vs. DragonFlyBSD 6 vs. Linux Message-ID: <b9b94d6d-93d0-4c3a-d07d-4f1d0cf90708@freebsd.org> In-Reply-To: <B53325BE-AEF5-4FD0-B791-93D3205CCEA4@FreeBSD.org> References: <CA%2BGLnbgVGghYAYPbQfu0H0cGvXxk-v0jAZTxLLz%2BhRn5eXjP0g@mail.gmail.com> <f8e569a4-3510-5a91-62b2-f9080b197ebc@beepc.ch> <CAJQ5JnhpioiO_j-iidk8QbfONU6a2Jm%2BunGCdAuN=Es8UZUQzw@mail.gmail.com> <B53325BE-AEF5-4FD0-B791-93D3205CCEA4@FreeBSD.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On 12/11/2021 5:38 AM, David Chisnall wrote: > While I agree on most of your points, the value of Phoronix is that it tests the default install. > > As an end user, I don’t care that a particular program is twice as fast on a particular Linux distro as it is on FreeBSD because of kernel features, compiler options, or dependency choices. > > I would love to see the base system include the ThinLTO (LLVM IR) .a files so that I can do inlining from libc into my program. I would love for ports to default to ThinLTO unless they break with it. Apple flipped that switch a few years ago, so a lot of things that broke with ThinLTO are now fixed. > > The FreeBSD memcpy / memset implementations look like they’re slower than the latest ones, which can give a 5-10% perf boost on some workloads. LLVM just landed the automemcpy framework, which is designed by some Google folks to synthesises efficient memcpy implementations tailored to different workloads. > > FreeBSD often wins versus glibc-based distros because jemalloc is faster than dlmalloc (the default malloc implementations in FreeBSD libc and glibc, respectively). I’ve been using snmalloc in my libc for a while and it generally gives me a few percent more perf. Unfortunately, FreeBSD decided to expose all of the jemalloc non-standard functions from libc, which means I can’t contribute it to upstream without implementing all of those on top of snmalloc or it would be an ABI break. > > It would be great if someone could pick up the Phronix benchmark suite and do some profiling: where is FreeBSD spending more time than Linux? Are there Linux-specific code paths that hit slow paths on FreeBSD and fast paths on Linux that could have FreeBSD-specific fast paths added (e.g. futex vs _umtx_op)? > > David > > >> On 11 Dec 2021, at 10:17, dmilith . <dmilith@gmail.com> wrote: >> >> 1. Where are compiler options for BSDs? >> 2. Why they compare -O2 to -O3 code in some benchmarks? Why they enable >> fast math in some, and disable it for others? >> 3. Why they don't mention powerd setup for FreeBSD? By default it may use >> slowest CPU mode. Did they even load cpufreq kernel module? >> 4. Did they even care about default FreeBSD mitigations (via sysctl) >> enabled, or it's only valid for Linuxes? ;) >> 5. What happened to security and environment details of BSDs? >> >> It's kinda known that guys from Phroenix lack basic knowledge of how to do >> proper performance testing and lack basic knowledge about BSD systems. >> Nothing new. Would take these results with a grain of salt. >> >> On Sat, 11 Dec 2021 at 10:53, beepc.ch <xpetrl@beepc.ch> wrote: >> >>>> I am surprised to see that the BSD cluster today has much worse >>> performance >>>> than Linux. >>>> What do you think of this? >>> >>> "Default" FreeBSD install setting are quite conservative. >>> The Linux common distros are high tuned, those benchmark is in my >>> opinion comparison of apples and oranges. >>> >>> Comparing "default" FreeBSD install with "default" Slackware install >>> would be more interesting, because Slackware builds are at most vanilla. >>> >>> >> >> -- >> Daniel Dettlaff >> Versatile Knowledge Systems >> verknowsys.com > > I tried to investigate this a bit yesterday. Specifically, seeing 'zstd' and 'openssl 3.0 sha256' being significantly worse on FreeBSD, when especially the latter, should basically have 0 difference between OSes. The Phoronix test suite is available as a package for freebsd, but I don't know what kind of magic is required to make it actually work. The compress-zstd-1.5.0 benchmark's install.sh tries to compile zstd with BSD make, when the makefile only works for GNU make. A little hacking around that, and I managed to run the test, but the results I got were in line with Linux, over 2 GB/sec, not the 300 MB/sec their result reported. I've not managed to figure out what they could have done wrong. It really does look like for zstd, it was someone only using 1 core on FreeBSD, and all cores on every other OS. -- Allan Jude
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?b9b94d6d-93d0-4c3a-d07d-4f1d0cf90708>