> On Sat, Oct 01, 2022 at 11:23:17PM -0000, Patchwork wrote: > > * igt@gem_exec_balancer@parallel-balancer: > > - shard-iclb: [PASS][58] -> [SKIP][59] ([i915#4525]) > > [58]: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_12204/shard- > iclb2/igt@gem_exec_balancer@xxxxxxxxxxxxxxxxxxxxxx > > [59]: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_109175v2/shard- > iclb5/igt@gem_exec_balancer@xxxxxxxxxxxxxxxxxxxxxx > > shard-iclb skips most DRRS tests, but does execute a few which is > weird. > > I spot checked a few of he logs and saw at least three different panels > being used: > 1. using preferred EDID fixed mode: "2560x1440": 60 241750 2560 2608 2640 > 2720 1440 1443 1448 1481 0x48 0xa > 2. using preferred EDID fixed mode: "1920x1080": 60 141000 1920 1936 1952 > 2104 1080 1083 1097 1116 0x48 0xa > 3. using preferred EDID fixed mode: "1920x1080": 60 138780 1920 1966 1996 > 2080 1080 1082 1086 1112 0x48 0xa > using alternate EDID fixed mode: "1920x1080": 40 92520 1920 1966 1996 2080 > 1080 1082 1086 1112 0x40 0xa > > So the DRRS tests were only executed when they ended up on machine with > panel 3. > > Having different panels between the machines in the shard pool is not > great. We can get all kinds of pingpongs depending on how tests get > scheduled to individual machines. ICL-shard is, sadly, heterogenous bunch. In addition to different panels, the CPUs themselves are different even if fusing is done to make them look almost the same. Considering that there are not many ICLs on market and we don't still have good choices for production CI hardware (as opposed to eg. TGL where we have both pre-prod and prod hardware readily available), we should consider taking shard-iclb out and leaving couple of them for BAT runs. Tomi
