On Mon, 14 Aug 2017 17:16:22 +1000, Aristedes Maniatis wrote: > On 14/8/17 3:08PM, Kevin Oberman wrote: > > Again, the documentation lags reality. The default was changed for > > 11.0. It is still conservative. In ALMOST all cases, Cmax will yield > > the bast results. However, on large systems with many cores, Cmax > > will trigger very poor results, so the default is C2, just to be > > safe. Given it's a server, anything beyond C2 is likely not worth trying. OTOH, C2 is perhaps not worth avoiding; it's probably low latency and should result in lower power consumption, so heat, and unlikely to hurt. Or at least, I suspect that's the case .. cc'ing Alexander, as the wiki article you referenced was his doing, so he's among those best placed. > > As far as possible TSC impact, I think older processors had TSC > > issues when not all cores ran with the same clock speed. That said, > > I am not remotely expert on such issues, so don't take this too > > seriously. I wasn't aware that FreeBSD could yet do different freqs on different cores? But I'm less expert than Kevin, and certainly behind the times. > Thanks Kevin > > What does 'large' and 'many cores' mean here? Is 24 cores large or > small? For a server do we ever want the CPU to enter states other > than C1? If C2 works well on your box, I don't see why you wouldn't want to use it .. but others might. I have no personal experience beyond 2 cores, but I'm perennially curious about such issues, as Kevin knows :) Are you using powerd? And what says, for example: sysctl -a | egrep 'cx|available|_freq|freq_|choice' | grep -v net\. which should include your timecounter and eventtimer setup too. cheers, Ian
On 14.08.2017 18:38, Ian Smith wrote:> On Mon, 14 Aug 2017 17:16:22 +1000, Aristedes Maniatis wrote: > > On 14/8/17 3:08PM, Kevin Oberman wrote: > > > Again, the documentation lags reality. The default was changed for > > > 11.0. It is still conservative. In ALMOST all cases, Cmax will yield > > > the bast results. However, on large systems with many cores, Cmax > > > will trigger very poor results, so the default is C2, just to be > > > safe. > > Given it's a server, anything beyond C2 is likely not worth trying. > OTOH, C2 is perhaps not worth avoiding; it's probably low latency and > should result in lower power consumption, so heat, and unlikely to hurt. > > Or at least, I suspect that's the case .. cc'ing Alexander, as the wiki > article you referenced was his doing, so he's among those best placed.C-states controlled here are ACPI C-states, which have limited relation to real CPU C-states. There are systems where they map exactly, but there are also systems where ACPI C1/C2/C3 states map to CPU C1/C3/C6, so it is difficult to make general recommendations. Approximately the map can be guessed looking on latency value (last of three) reported in sysctl dev.cpu.0.cx_supported: 1 is usually CPU C1, 2+ is likely CPU C2, 100+ can be C3, 500+ can be C6, but all that is very approximately and I guess depends on BIOS writer mood. What's about recommendations from me, I'd say that CPU C2 state should not hurt in most cases, unless something is broken, but benefit is rather small (often just covered by C1E enabled in BIOS); CPU C3 state gives significant power saving, but can either hurt performance due to higher enter/exit latency or slightly improve it due to TurboBoost activation (require CPU frequency to be set to max value); CPU C6 is probably useful only for laptops, since it saves not so much power power, while exit latency can be in milliseconds range.> > > As far as possible TSC impact, I think older processors had TSC > > > issues when not all cores ran with the same clock speed. That said, > > > I am not remotely expert on such issues, so don't take this too > > > seriously. > > I wasn't aware that FreeBSD could yet do different freqs on different > cores? But I'm less expert than Kevin, and certainly behind the times.On old CPUs TSC frequency was related to CPU frequency and so could fluctuate with frequency change. On modern CPUs it is always constant, equal to base CPU frequency. What's about different frequency for different cores, IIRC ACPI allows that, but up to recent time neither FreeBSD nor hardware could do that. I have feeling I heard that some very new CPUs may allow that, but to be efficient it would require very tight interoperation between power manager and CPU scheduler, otherwise performance may suffer. -- Alexander Motin
On Mon, Aug 14, 2017 at 8:38 AM, Ian Smith <smithi at nimnet.asn.au> wrote:> On Mon, 14 Aug 2017 17:16:22 +1000, Aristedes Maniatis wrote: > > > On 14/8/17 3:08PM, Kevin Oberman wrote: > > > Again, the documentation lags reality. The default was changed for > > > 11.0. It is still conservative. In ALMOST all cases, Cmax will yield > > > the bast results. However, on large systems with many cores, Cmax > > > will trigger very poor results, so the default is C2, just to be > > > safe. > > Given it's a server, anything beyond C2 is likely not worth trying. > OTOH, C2 is perhaps not worth avoiding; it's probably low latency and > should result in lower power consumption, so heat, and unlikely to hurt. > > Or at least, I suspect that's the case .. cc'ing Alexander, as the wiki > article you referenced was his doing, so he's among those best placed. > > > > As far as possible TSC impact, I think older processors had TSC > > > issues when not all cores ran with the same clock speed. That said, > > > I am not remotely expert on such issues, so don't take this too > > > seriously. > > I wasn't aware that FreeBSD could yet do different freqs on different > cores? But I'm less expert than Kevin, and certainly behind the times. > > > Thanks Kevin > > > > What does 'large' and 'many cores' mean here? Is 24 cores large or > > small? For a server do we ever want the CPU to enter states other > > than C1? > > If C2 works well on your box, I don't see why you wouldn't want to use > it .. but others might. I have no personal experience beyond 2 cores, > but I'm perennially curious about such issues, as Kevin knows :) > > Are you using powerd? And what says, for example: > > sysctl -a | egrep 'cx|available|_freq|freq_|choice' | grep -v net\. > > which should include your timecounter and eventtimer setup too. > > cheers, Ian >I guess I need to clarify. No, FreeBSD does not have the ability to tun different cores at different frequencies. I seem to recall that TCC on some processors could adjust the frequency of a core exceeding a defined temperature, skipping N of every 8 clock cycles to slow the processor and reduce the temperature. This is what TCC was designed for. It is entirely possible that I am not correctly remembering the details of the issue, but it could only be resolved by switching from TCC to another clocking system. If memory serves, and it may not, there was an issue a few years ago (jhb@ worked the issue) where TSC was varying with frequency and that caused clock drift. I believe all "modern" processors do not have this issue and it seems unlikely that any system running 24 cores is old enough that this might be an issue. Sorry for any confusion I may have caused.