Date: Wed, 22 Mar 2023 21:09:22 -0700 From: Mark Millard <marklmi@yahoo.com> To: George Mitchell <george+freebsd@m5p.com> Cc: FreeBSD Hackers <freebsd-hackers@freebsd.org> Subject: Re: Periodic rant about SCHED_ULE Message-ID: <6822D3CE-789E-4FB0-BB62-BEEEA65DB72F@yahoo.com> In-Reply-To: <F9A9FCBF-C46A-455C-A18F-277521BFF2A4@yahoo.com> References: <6BD317F2-7EDD-45C0-9DC9-5B94C1BBB8E1.ref@yahoo.com> <6BD317F2-7EDD-45C0-9DC9-5B94C1BBB8E1@yahoo.com> <952d9795-19dc-8ad1-bb75-5c556ca6795a@m5p.com> <E8A93C10-F5A4-4473-9AED-299C108CAD6C@yahoo.com> <78EF511D-BAF1-495F-BAC9-03AC1B8FD56A@yahoo.com> <A0F19E2A-A63E-4CBD-B962-CEEC455BBC0A@yahoo.com> <E30D9325-60C4-42A8-9929-5F51F8F102E0@yahoo.com> <0196A767-5DC2-4DA3-BA5E-62A358CBCF64@yahoo.com> <F9A9FCBF-C46A-455C-A18F-277521BFF2A4@yahoo.com>
next in thread | previous in thread | raw e-mail | index | archive | help
[I added a -j32 buildworld buildkernel with SCHED_4BSD and dnetc-in-use comparison, to the other ThreadRipper 1950X examples. SCHED_4BSD does take notably less time than SCHED_ULE when dnetc is also active: still a good match to the simple round-robin for this building activity. I will note that the 1950X UEFI/firmware is not configured present itself as NUMA but the FreeBSD kernels in use are NUMA capable as built.] On Mar 22, 2023, at 19:44, Mark Millard <marklmi@yahoo.com> wrote: > On Mar 22, 2023, at 18:08, Mark Millard <marklmi@yahoo.com> wrote: >=20 >> On Mar 22, 2023, at 18:03, Mark Millard <marklmi@yahoo.com> wrote: >>=20 >>> On Mar 22, 2023, at 16:17, Mark Millard <marklmi@yahoo.com> wrote: >>>=20 >>>> On Mar 22, 2023, at 15:39, Mark Millard <marklmi@yahoo.com> wrote: >>>>=20 >>>>> On Mar 22, 2023, at 13:34, Mark Millard <marklmi@yahoo.com> wrote: >>>>>=20 >>>>>> On Mar 22, 2023, at 12:40, George Mitchell = <george+freebsd@m5p.com> wrote: >>>>>>=20 >>>>>>> On 3/22/23 15:21, Mark Millard wrote: >>>>>>>> George Mitchell <george+freebsd@m5p.com> wrote on >>>>>>>> Date: Wed, 22 Mar 2023 17:36:39 UTC : >>>>>>>> [...] >>>>>>>>> Here are the very complicated instructions for reproducing the = problem: >>>>>>>>> 1. Install and start misc/dnetc from ports. >>>>>>>> Installing is likely easy, as likely would be building >>>>>>>> with default options (if any). I know nothing about >>>>>>>> starting misc/dnetc so that is research. (Possibly >>>>>>>> trivial, although if it has alternatives to control >>>>>>>> then I'd need to match that context too.) >>>>>>>=20 >>>>>>> service dnetc start >>>>>>=20 >>>>>> I built and installed misc/dnetc and got a binary >>>>>> blob that clearly was not built in my environment: >>>>>>=20 >>>>>> # file /usr/local/distributed.net/dnetc >>>>>> /usr/local/distributed.net/dnetc: ELF 64-bit LSB executable, = x86-64, version 1 (FreeBSD), statically linked, for FreeBSD 10.1 = (1001515), FreeBSD-style, stripped >>>>>>=20 >>>>>> Way older FreeBSD vintage than the locally available toolchains >>>>>> would normally build. Some might be cautious about such a thing. >>>>>>=20 >>>>>> The man page reported that: >>>>>>=20 >>>>>> QUOTE >>>>>> If you have never run the client before, it will initiate the = menu-driven >>>>>> configuration. Save and quit when done, the configuration file = will be >>>>>> saved in the same directory as the client. Now, simply restart = the >>>>>> client. =46rom that point on it will use the saved configuration. >>>>>> END QUOTE >>>>>>=20 >>>>>> I've not seen what the configuration asks about yet. >>>>>=20 >>>>> I went through the configuration, basically just looking >>>>> at it, other than providing an E-mail address. Then . . . >>>>>=20 >>>>> $ sudo service dnetc start >>>>> Password: >>>>> Cannot 'start' dnetc. Set dnetc_enable to YES in /etc/rc.conf or = use 'onestart' instead of 'start'. >>>>>=20 >>>>> $ sudo service dnetc onestart >>>>>=20 >>>>> I just let it run without any extra competing activity, other >>>>> than I had my patched version of top running. It records and >>>>> reports various maximum-observed (MaxObs) figures, here >>>>> the load averages being relevant. >>>>>=20 >>>>> Top showed that dnetc started 32 processes, one per hardware >>>>> thread. Mostly I saw: 100% nice and 0% idle. >>>>>=20 >>>>> Letting it run and then looking at the load averages (and >>>>> their matching MaxObs figures) after something like 60+ min >>>>> (not carefully timed: was doing other things) showed: >>>>>=20 >>>>> load averages: 31.97, 31.88, 31.66 MaxObs: 32.12, 31.97, = 31.66 >>>>>=20 >>>>> (Note: The machine had been up for over 2.75 days before >>>>> starting this and had not been building much of anything >>>>> during that time.) >>>>>=20 >>>>> I've not yet experimented with having other, significant >>>>> competing activity. >>>>>=20 >>>>>>>>> 2. Run "make buildworld". >>>>>>>> So on the 32 hardware-thread (16 cores) amd64 machine that >>>>>>>> I have access to, the test is to only have buildworld use >>>>>>>> about one hardware thread, no matter what else is going on. >>>>>>>> I never would have guessed that the steps would not involve >>>>>>>> more like -j$(sysctl -n hw.ncpu) (so around -j32 in this >>>>>>>> context). So it is good that you provided your note or >>>>>>>> I'd not know if I'd done similarly or not when trying such. >>>>>>>> [Note: -j1 and lack of -j are not strictly equivalent in >>>>>>>> how make operates. As I remember, the distinction makes >>>>>>>> a notable difference in the number of subprocesses created >>>>>>>> directly by make (one per action "line" vs. one for the >>>>>>>> whole block?). So even using -j1 might make a difference >>>>>>>> vs. what you specified. I'd have to test to see.] >>>>>>>=20 >>>>>>> I am literally running "make buildworld" with no additional = options. >>>>>>=20 >>>>>> So required for repeating your results, but likely making >>>>>> such results not be interesting relative to how I normally >>>>>> deal with buildworld buildkernel and the likel, no matter >>>>>> if there is other activity in an overlapping time frame or >>>>>> not: my time preferences are too strong to wait for a single >>>>>> hardware thread to do my normal builds, even with no >>>>>> competing activity on the builder. >>>>>>=20 >>>>>>>>> Standard out conveniently reports how long it took (wall = clock). >>>>>>>> But nothing in your instructions indicate about how >>>>>>>> to get an idea much progress dnetc made during the >>>>>>>> various tests? [...] >>>>>>>=20 >>>>>>> Honestly, I've never worried about this part. But dnetc logs = its >>>>>>> progress in /usr/local/distributed.net/dnetc.txt, though not in = terms >>>>>>> that are easy to relate to real-world progress. Oddly, when I = run >>>>>>> "make buildworld," I'm primarily interested in getting the world = built. >>>>>>> Perhaps others feel differently. >>>>>>=20 >>>>>> Off topic for the specifics of the actual benchmark >>>>>> that you run: >>>>>>=20 >>>>>> Then why not use of -jN ? In my context, any buildworld >>>>>> using -j1 or no -j at all takes a huge amount of time >>>>>> longer than letting it use all the hardware threads (or >>>>>> so). (I've avoided having any I/O bound contexts for >>>>>> such.) It does not take additional load on the system >>>>>> for that to be true --including on the 4-core small arm >>>>>> boards when I happen to buildworld on such (rare). >>>>>>=20 >>>>>>=20 >>>>>>>> [...] >>>>>>>> FYI: I've never built with and run the alternate >>>>>>>> scheduler so if there is any appropriate background >>>>>>>> for that that would not be obvious on finding basic >>>>>>>> instructions, it would be appropriate to provide >>>>>>>> such notes. >>>>>>>> [...] >>>>>>>=20 >>>>>>> You have to build a new kernel, using a config file in which you = have >>>>>>> replaced "options SCHED_ULE" with "options SCHED_4BSD". -- = George >>>>>>=20 >>>>>> Thanks for the notes. >>>>>>=20 >>>>>> I've not decided if I'll do anything with the binary >>>>>> blob or not. >>>>>=20 >>>>=20 >>>> FYI: >>>>=20 >>>> It is not your specific experiment, but I started my >>>> "extra load" experimenst with . . . >>>>=20 >>>> I started a -j32 buildworld buildkernel with dnetc still >>>> running. I'm generally seeing around 55% Active and 42% >>>=20 >>> Note "Active": user, sorry. >>>=20 >>>> nice, < 2% system (it was building libllvm at this point). >>>> At that time: >>>>=20 >>>> load averages: 64.41, 60.52, 49.81 MaxObs: 64.47, 60.52, = 49.81 >>>>=20 >>>=20 >>> Contrasting results for some obj-lib32 build activity: >>> much more variety of User, nice, and system, including >>> times with < 5% user, 90+% nice. But not typical overall. >>> But lots of time roughly around 50%/50% or 35%/60%. There >>> were times with 15+% system. >>>=20 >>> Somewhat after buildkernel started: >>>=20 >>> load averages: 69.15, 64.12, 58.72 MaxObs: 75.98, 64.12, 58.72 >>>=20 >>> Harder to summarize, so overall timing reports from the >>> buildworld and buildkernel stages. >>>=20 >>>=20 >>> buildworld: >>>=20 >>> -------------------------------------------------------------- >>> ... World build completed on Wed Mar 22 16:37:57 PDT 2023 >>> ... World built in 2615 seconds, ncpu: 32, make -j32 >>> -------------------------------------------------------------- >>>=20 >>>=20 >>> buildkernel: >>>=20 >>> -------------------------------------------------------------- >>> ... Kernel build for GENERIC-NODBG completed on Wed Mar 22 16:43:10 = PDT 2023 >>> -------------------------------------------------------------- >>> ... Kernel(s) GENERIC-NODBG built in 311 seconds, ncpu: 32, make = -j32 >>> -------------------------------------------------------------- >>>=20 >>> Afterwards: >>>=20 >>> load averages: 36.08, 53.14, 55.79 MaxObs: 75.98, 65.77, 59.84 >>>=20 >>>=20 >>> I then did (not all in the same window): >>>=20 >>> $ sudo service dnetc onestop >>> # rm -fr /usr/obj/BUILDs/main-amd64-nodbg-clang-alt/usr/ >>>=20 >>> before another -j32 buildworld buildkernel (no dnetc). The >>> reuslts for this were: >>>=20 >>>=20 >>> buildworld: >>>=20 >>> -------------------------------------------------------------- >>> ... World build completed on Wed Mar 22 17:39:19 PDT 2023 >>> ... World built in 1240 seconds, ncpu: 32, make -j32 >>> -------------------------------------------------------------- >>>=20 >>> (compared to the 2615 for dnetc also in use) >>>=20 >>>=20 >>> buildkernel: >>>=20 >>> -------------------------------------------------------------- >>> ... Kernel build for GENERIC-NODBG completed on Wed Mar 22 17:41:17 = PDT 2023 >>> -------------------------------------------------------------- >>> ... Kernel(s) GENERIC-NODBG built in 118 seconds, ncpu: 32, make = -j32 >>> -------------------------------------------------------------- >>>=20 >>> (compared to the 311 for dnetc also in use) >>=20 >> I forgot to show the MaxObs load averages for the no-dnetc >> context: >>=20 >> MaxObs: 39.77, 32.15, 25.75 >>=20 >>> Experiments without -j32 will take a lot longer, even >>> without dnetc in use. I'm not sure there will be such >>> results today. >>>=20 >>=20 >=20 > I decided to do some more of the less time consuming > testing. SCHED_4BSD, no dnetc, -j32 buildworld buildkernel : >=20 >=20 > buildworld: >=20 > -------------------------------------------------------------- > ... World build completed on Wed Mar 22 19:16:35 PDT 2023 > ... World built in 1235 seconds, ncpu: 32, make -j32 > -------------------------------------------------------------- >=20 > (compared to 1240 for SCHED_ULE) >=20 > So: no significant difference. >=20 >=20 > buildkernel (SCHED_4BSD building a SCHED_4BSD): >=20 > -------------------------------------------------------------- > ... Kernel build for GENERIC-NODBG-SCHED_4BSD completed on Wed Mar 22 = 19:18:34 PDT 2023 > -------------------------------------------------------------- > ... Kernel(s) GENERIC-NODBG-SCHED_4BSD built in 119 seconds, ncpu: = 32, make -j32 > -------------------------------------------------------------- >=20 > (compared to 118 for SCHED_ULE building a SCHED_ULE) >=20 > So: no significant difference. I again forgot to show MaxObs load averages (for the above): MaxObs: 39.23, 31.58, 24.30 > I'll try it with dnetc also active. >=20 I still have no good indication of dnetc progress to allow comparison of the combination. So the below focuses on buildworld buildkernel . I expect that the comparative results suggest a buildworld/buildkernel vs. dnetc progress tradeoff, not that I can well quantify it. The below are with dnetc also active. load averages, MaxObs: 73.03, 65.48, 56.30 (I remembered this time!) buildworld: -------------------------------------------------------------- ... World build completed on Wed Mar 22 20:15:56 PDT 2023 ... World built in 1667 seconds, ncpu: 32, make -j32 -------------------------------------------------------------- (compared to 2615 for SCHED_ULE with dnetc and to 1240 or so for no dnetc) buildkernel: -------------------------------------------------------------- ... Kernel build for GENERIC-NODBG-SCHED_4BSD completed on Wed Mar 22 = 20:18:34 PDT 2023 -------------------------------------------------------------- ... Kernel(s) GENERIC-NODBG-SCHED_4BSD built in 158 seconds, ncpu: 32, = make -j32 -------------------------------------------------------------- (compared to 311 for SCHED_ULE with dnetc and to 118 or so for no dnetc) With dnetc active, it does not take being near -j1 (or no -j) for buildworld buildkernel to take noticably less time: -j32 (the number of hardware threads, 16 cores) also takes noticeably less time. buildworld buildkernel in this context seems to be a good match to SCHED_4BSD and its round-robin. (I make no general claim to SCHED_4BSD being better across a large range of contexts.) I've not decided if I'll try anything like a -j1 or no -j alternative. Without dnetc active, SCHED_ULE and SCHED_4BSD did not make much of a distinction. For how I use the builder machines, the scheduler choice is not suggested to be significant for my system-build activities. I've not tested port building in poudriere-devel for how I configure such. But nothing suggests to me to expect a significant distinction between the 2 schedulers for my way of working for building packages from ports. =3D=3D=3D Mark Millard marklmi at yahoo.com
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?6822D3CE-789E-4FB0-BB62-BEEEA65DB72F>