== Week of March 24th ==
- STREAM regression (TCWG-388, 5/10) -- Finished prototype patch. The patch adds modeling of ARM L2 auto-prefetcher hardware to GCC scheduler (the model is very simple as auto-prefetcher is very lightly documented). Half of the patch cleans up and improves GCC scheduler, and the other half implements the auto-prefetcher model. -- While looking into ARM scheduling support noticed that ARM doesn't use multipass lookahead scheduling, which surprised me. Enabled it (multipass scheduling) in my patches.
- Looked into lll_timed_wait Glibc/uClibc bug upstream (1/10) -- https://sourceware.org/ml/libc-alpha/2014-03/msg00905.html
- Various discussions and reviews (4/10)
== Week of March 30th ==
- STREAM regression (TCWG-388) -- Benchmark patches on SPEC2k and find/confirm best values for tuning parameters: --- dfa_lookahead: should normally be issue_rate-1. --- L2 auto-prefetcher queue depth: new tuning knob. -- Investigate any performance regressions from the patches.
- lll_timed_wait Glibc/uClibc bug -- Make sure it is fixed upstream. Possibly backport to Linaro branches.
-- Maxim Kuvyrkov www.linaro.org
On Apr 1, 2014, at 5:48 AM, Maxim Kuvyrkov maxim.kuvyrkov@linaro.org wrote:
== Week of March 24th ==
- STREAM regression (TCWG-388, 5/10)
-- Finished prototype patch. The patch adds modeling of ARM L2 auto-prefetcher hardware to GCC scheduler (the model is very simple as auto-prefetcher is very lightly documented). Half of the patch cleans up and improves GCC scheduler, and the other half implements the auto-prefetcher model. -- While looking into ARM scheduling support noticed that ARM doesn't use multipass lookahead scheduling, which surprised me. Enabled it (multipass scheduling) in my patches.
- Looked into lll_timed_wait Glibc/uClibc bug upstream (1/10)
-- https://sourceware.org/ml/libc-alpha/2014-03/msg00905.html
- Various discussions and reviews (4/10)
== Week of March 30th ==
- STREAM regression (TCWG-388)
-- Benchmark patches on SPEC2k and find/confirm best values for tuning parameters: --- dfa_lookahead: should normally be issue_rate-1. --- L2 auto-prefetcher queue depth: new tuning knob. -- Investigate any performance regressions from the patches.
Almost forgot:
In preparation for benchmarking runs figured out how to configure chromeos+crouton chromebook for benchmarking (this avoids background Chrome processes from clobbering performance results): 1. At the login screen pressing ctrl+alt+f2 to switch to VT2 2. login into shell at VT2 3. "sudo stop ui" will stop session_manager and kill all chrome processes 4. switch back to VT2 (ctrl+alt+f2) 5. enter crouton chroot: "sudo startcli" 6. ssh to your chroot 7. run cpufreq to run CPUs at full throttle 8. run benchmarks The caveat is that only CLI-type chroot is suitable with chrome-os's X server down.
Additionally, I've added NFSv4 support to said crouton chroot by rebuilding modules for ChromeOS kernel with NFSv4 included (https://github.com/dnschneid/crouton/wiki/Build-chrome-os-kernel-and-kernel-...).
- lll_timed_wait Glibc/uClibc bug
-- Make sure it is fixed upstream. Possibly backport to Linaro branches.
-- Maxim Kuvyrkov www.linaro.org
-- Maxim Kuvyrkov www.linaro.org
linaro-toolchain@lists.linaro.org