On 19/08/16 15:17, Henrique de Moraes Holschuh wrote:
It means 0x36 can be safely upgraded to 0x38, and 0x36 is widely
available and used, so it is valuable info.
Ok, I must have misinterpreted your previous question. No matter.
I down-graded the BIOS, and spent a while alternating between microcodes
0x32 (as shipped in the BIOS), 0x36 (intel-microcode_3.20151106) and
0x38 (intel-microcode_3.20160714). In combination with this BIOS
version, 0x36 was previously problematic.
I reproducibly found that on initial booting, with no system load, 0x36
came up with processors clocked at 900MHz (below the cpufreq
scaling_min_freq of 1200MHz), before then settling to 1200MHz after some
time, whereas 0x32 and 0x38 behaved alike, with the cores at much higher
initial frequencies (later settling to 1200MHz).
I have spent a few days since trying to reliably reproduce the `stuck at
400MHz' behaviour with 0x36, but the server's workload has changed since
March, and the 400MHz behaviour is less frequent -- I've seen it twice.
Using the 0x38 microcode I have not yet had any 400MHz occurrences BUT
the time 0x38 has been on test is much less.
So, in summary, the initial signs are good that 0x38 may fix the 0x36
performance regression, but I think a couple more weeks' use would make
me more certain of that.
Stuart
