| === JanC is now known as Guest79577 | ||
| === JanC_ is now known as JanC | ||
| cyphermox | fyi; https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1673350 | 01:50 |
|---|---|---|
| ubot5 | Ubuntu bug 1673350 in multipath-tools (Ubuntu) "dm-queue-length module is not included in installer/initramfs" [Critical,Triaged] | 01:50 |
| cyphermox | not sure who to assign this to, but I would need dm-queue-length added to multipath-modules | 01:51 |
| smb | cyphermox, I thought there was some talk about that for s390x... | 09:02 |
| smb | cyphermox, ok, I assigned myself and added yakkety. likely should do the udeb changes there as well | 09:07 |
| xnox | smb, there was this s390x ticket https://bugs.launchpad.net/ubuntu/+source/multipath-tools/+bug/1634161 | 10:29 |
| ubot5 | Ubuntu bug 1634161 in multipath-tools (Ubuntu) "Default multipath policy is round-robin, shows considerably worse read throughput compared to using service-time" [Low,Confirmed] | 10:29 |
| xnox | smb, if you can include service time modules into udeb that would be nice for that ticket. | 10:29 |
| smb | xnox, Ah could not remember the number but yes service-time now is in the list and I just sent a patch to add queue-length as well | 10:30 |
| smb | That should be all known ones | 10:30 |
| smb | xnox, so service-time was added since xenial (or should be, I did not look at the udeb, only the file that should create it) | 10:32 |
| xnox | ack, let me look at the build-log and the .udeb contents files. one sec. | 10:34 |
| smb | xnox, -rw-r--r-- root/root 10158 2017-03-08 20:06 ./lib/modules/4.4.0-67-generic/kernel/drivers/md/dm-service-time.ko | 10:36 |
| xnox | yeap looking sexy. | 10:36 |
| xnox | cyphermox, would you be able to change defaults to service-time? https://bugs.launchpad.net/ubuntu/+source/multipath-tools/+bug/1634161 | 10:38 |
| ubot5 | Ubuntu bug 1634161 in multipath-tools (Ubuntu) "Default multipath policy is round-robin, shows considerably worse read throughput compared to using service-time" [Low,Confirmed] | 10:38 |
| xnox | in multipath-tools, or shall i? | 10:38 |
| * xnox is not sure if we should be doing this this close to release..... | 10:38 | |
| apw | xnox, i doubt anyone will have specific testing before now using the existing default, it is a pretty nieche thing | 10:48 |
| xnox | apw, what do you mean? as far as I understand upstream multipath-tools default is service-time; yet we switch to round-robin. And there is some tests/benchmarks done in the bug report indicating that service-time is better. | 10:51 |
| xnox | so which testing are you referring to? | 10:51 |
| xnox | as in testing service-time to make sure it is better? | 10:52 |
| smb | xnox, apw I would think that the mainframe people have a good history of using dm-queue-length (initially at least) or dm-service-time. Those both would generally be better handling variations in IO time on different paths. While round-robin dumbly switches every 1000 requests | 10:58 |
| apw | or testing it either way | 10:58 |
| apw | i mean changing is not so bad imo | 10:59 |
| smb | Yeah I would not really expect too horrific issues | 11:01 |
| smb | IIRC it should be possible to change the balancer even post-installation via multipath.conf... | 11:04 |
| xnox | correct. but we should use sane default | 11:41 |
| cyphermox | xnox: I'll look at it, there's another fix I need to do in multipath-tools anyway | 13:13 |
| cyphermox | ah, I see | 13:15 |
| cyphermox | yeah, it was round-robin only because service-time wasn't available in wily | 13:16 |
| xnox | i think consensus above it to just drop the patch; such that we use the upstream default of service-time by default. it is in udebs at least in xenial-release and up. | 13:16 |
| cyphermox | yup | 13:18 |
| cyphermox | the fact that it's still round-robin instead is a bug | 13:19 |
| cyphermox | I'll fix this while also fixing queue-length | 13:19 |
| cyphermox | smb: does dm_emc still exist anywhere? | 13:20 |
| cyphermox | (IIRC it doesn't, but trying to make sure) | 13:21 |
| smb | cyphermox, hm... don't think so... though maybe you mean the device handler... give me a sec | 13:21 |
| smb | cyphermox, scsi_dh_emc.ko | 13:22 |
| cyphermox | nah, really "dm-emc" | 13:23 |
| cyphermox | it used to be a module, and I think I removed it in d-i already | 13:23 |
| smb | cyphermox, cannot even remember that | 13:23 |
| cyphermox | yep | 13:23 |
| cyphermox | - Remove dm-emc from the multipath modules, since it's gone since | 13:23 |
| cyphermox | 2.6.27. | 13:23 |
| cyphermox | if I didn't lie when I wrote that | 13:23 |
| smb | heh | 13:24 |
| smb | cyphermox, though I really think that if dm-emc was multipath related then it likely was what scsi-dh-emc is now (and those are at least in scsi modules) | 13:28 |
| cyphermox | yup, I agree | 13:28 |
| cyphermox | the SRU for service-time is going to be yucky | 13:36 |
| cyphermox | xnox: fyi, will upload in a minute just being careful and building locally first. | 13:41 |
| xnox | cyphermox, why would SRU for service-time be yucky? no change to kernel or udebs.... just a flip of default. | 13:42 |
| cyphermox | yeah, a flip of default which will flip running systems. | 13:50 |
| cyphermox | ie. you could get better throughput or explosions. | 13:50 |
| === cachio is now known as cachio_afk | ||
Generated by irclog2html.py 2.7 by Marius Gedminas - find it at mg.pov.lt!