On Jul 17, 2017, at 10:48 AM, Christophe Lyon christophe.lyon@linaro.org wrote:

• Short week (Friday off)

== Progress ==

• Infrastructure/validation:

• Benchmarking:

• production scripts are now up-to-date, still observing noise on

some benchmarks

I've investigate noise in 401.bzip2 (6%) and 436.cactusADM (10%). In both cases the extra cycle counts were on memory instructions, which suggests different memory or cache layout.

One of the things we can try is to reboot boards before high-accuracy runs to, hopefully, get roughly the same physical addresses from the kernel.

We also should try adding cache-misses counters to perf sampling to confirm/deny our assumptions about memory subsystem.

• fixed a couple of small issues with the scripts

• GCC upstream validation:

• working on further reducing noise ("random" pass/failures)
• reported/fixed a few regressions
• committed testsuite patch with de-require-stack-check
• looking at a problem with unaligned accesses

• misc (conf-calls, meetings, emails, ....)

-- Maxim Kuvyrkov www.linaro.org