> On June 1, 2016, 3:09 p.m., Qian Zhang wrote: > > src/slave/containerizer/mesos/containerizer.cpp, line 150 > > <https://reviews.apache.org/r/45086/diff/13/?file=1347681#file1347681line150> > > > > Not sure why you want to add `cgroups/cpuacct` here. Actually in the > > `creator` hashmap below, I do not see `cgroups/cpuacc` as a key, i.e., no > > create() method corresponding to `cgroups/cpuacc`. So I guess an `Unknown > > or unsupported isolator` error will happen when `cgroups/cpuacc` is > > specified as the value of `--isolation`. > > haosdent huang wrote: > Hi, @qianzhang Thanks a lot for your review! We need add cgroups/cpuacct > here because we split the original cpu isolator which contains cpu and > cpuacct into two subsystems. To compatiable old behaviour, we add it back at > here. > > Qian Zhang wrote: > @haosdent, you are welcome! I agree this is the right way to keep the > backward compatibility. However, as I mentioned in my comment above, the > question is in the `creator` hashmap, there is no create() method for > `"cgroups/cpuacc"`, that means when agent is started with > `"--isolation=cgroups/cpuacct"` or `"--isolation=cgroups"`, we will not find > the related create() method, so the `"Unknown or unsupported isolator"` error > will occur.
Oh, got it. Yes, If `cgroups/cpuacct` comes first in `--isolation` flag, it would return error, let me fix this. > On June 1, 2016, 3:09 p.m., Qian Zhang wrote: > > src/slave/containerizer/mesos/containerizer.cpp, lines 242-245 > > <https://reviews.apache.org/r/45086/diff/13/?file=1347681#file1347681line242> > > > > This seems not consistent with the design doc. In the design doc, I > > think we are going to introduce a new isolation mechanism `cgroups/unified` > > and a new flag `--cgroups_subsystems`, but here I see we reuse the existing > > cgroups isolation mechanisms. Maybe the design doc should be updated? > > haosdent huang wrote: > Yes, the design doc need to be updated. We would like to add this without > any changes when user upgrade. > > Qian Zhang wrote: > Understood, actually I also like the idea to introduce this new feature > without any changes when user upgrades. However I am thinking how we are > going to this case during upgrade: In an old Mesos cluster, user has enabled > `"cgroups/cpu"` for the agent which means internally both `cpu` and `cpuacct` > subsystems have been enabled, however after upgrading the cluster with this > new feature with the same flags to launch the agent, only `cpu` subsystem > will be enabled, right? So after upgrade we loose one subsystem (`cpuacct`)? Yes, we need to highlight this in upgrade document. - haosdent ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/45086/#review135801 ----------------------------------------------------------- On April 16, 2016, 10:29 a.m., haosdent huang wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/45086/ > ----------------------------------------------------------- > > (Updated April 16, 2016, 10:29 a.m.) > > > Review request for mesos, Gilbert Song, Guangya Liu, Alex Clemmer, Ian > Downes, Jie Yu, and Kevin Klues. > > > Bugs: MESOS-5041 > https://issues.apache.org/jira/browse/MESOS-5041 > > > Repository: mesos > > > Description > ------- > > Enable cgroups unified isolator in isolation. > > > Diffs > ----- > > src/slave/containerizer/mesos/containerizer.cpp > 1e1a36903f4377497bb72b69e4ead63675d453c0 > > Diff: https://reviews.apache.org/r/45086/diff/ > > > Testing > ------- > > > Thanks, > > haosdent huang > >
