To be honest, I don't know how exactly tracking this. What I am doing: Open 1 terminal window which runs latency test. With second terminal window I watch -n1 cat /proc/interrupts
During idle (beside latency test and watching interrupts), there are about 20 MSI / sec. If another window (e.g. File Manger) is moved around fast, I see about 500 MSI / sec. With Haswell, there are no i915 MSI Thanks for the hint about autotune. I will check with next measurements. -- Martin. -----Ursprüngliche Nachricht----- Von: Philippe Gerum [mailto:r...@xenomai.org] Gesendet: Mittwoch, 14. März 2018 15:31 An: Schrenkhammer, Martin; xenomai@xenomai.org Betreff: Re: [Xenomai] Latency Peaks with Skylake vs. Haswell On 03/13/2018 12:33 PM, Schrenkhammer, Martin wrote: > I am currently porting a project from Haswell to Skylake (i5-6440EQ). When in > desktop, I observe latency peaks when windows are moved around or zoomed. > With Haswell CPUs, there is not impact of graphic workload. Unfortunately > everything above 10us is an issue in our real time system application. > Currently I am using Xubuntu 16.04 + Kernel 4.9.51 + Xenomai 3.0.6 I > already disabled all power- and performance features like > Hyper-Threading, SpeedStep, Turbo Mode, C-States in BIOS and also by > kernel parameters. Also disabled i915 power savings like > i915.enable_dc=0, i915.enable_rc6=0, i915.enable_execlists=0... and > disabled grafik turbo gt_max_freq_mhz = 350 gt_boost_freq_mhz = 350 > > One difference we see between Haswell and Skylake: i915 raises a lot of MSI > when moving windows around where there are almost zero at Haswell. > > Any ideas for bringing the peaks down? Did you track which specific event the i915 driver was handling with those MSIs? > > == Sampling period: 100 us > == Test mode: periodic user-mode task > == All results in microseconds > warming up... > RTT| 00:00:01 (periodic user-mode task, 100 us period, priority 99) > RTH|----lat min|----lat avg|----lat max|-overrun|---msw|---lat > RTH|best|--lat worst > RTD| -0.368| -0.210| 0.259| 0| 0| -0.368| > 0.259 > RTD| -0.259| -0.161| 1.601| 0| 0| -0.368| > 1.601 > RTD| -0.290| -0.174| 1.367| 0| 0| -0.368| > 1.601 This is unrelated to the issue at stake, but the core timer looks slightly early (negative mins), which means that the expected intrinsic latency value is too pessimistic/high. Running autotune may help in picking better gravity values for this timer. -- Philippe. Vertrauliche E-Mail von / Confidential e-mail from: DATA MODUL AG Vorstand / CEO: Dr. Florian Pesahl Vorsitzende des Aufsichtsrates / Chairwoman of the Supervisory Board: Kristin D. Russell Sitz der Gesellschaft / Registered Office: München Registergericht / Registration Court: München Handelsregister B 85 591 _______________________________________________ Xenomai mailing list Xenomai@xenomai.org https://xenomai.org/mailman/listinfo/xenomai
