It's GiB, but there is a slice of per-task resource overhead added for the executor. I believe the default is 128 MB RAM for that, which seems to line up with what you are seeing.
On Wednesday, December 9, 2015, Andrew Jorgensen <[email protected]> wrote: > I am currently using aurora version 0.9.0-rc0-2 and mesos 0.23.0 > > I have a job that's stuck pending waiting for resources, specifically it > is waiting for ram. The job config says that the job needs 3 cores, 7.00 > GiB ram, and 16GiB disk. If I look at the mesos offers page I can see > that there are a number of machines that have 7.1GB available and if I > look at the offers in aurora I see a number of nodes that look like: > > {..., > > "hostname":"host","resources":[{"name":"cpus","scalar":7.0},{"name":"mem","scalar":7236.0},{"name":"disk","scalar":253167.0, > ...} > > So it looks to me that there is enough memory to schedule the pending > tasks but they are not actually getting scheduled. So one question is > whether the mesos ui and the aurora ui are using the same units > (gibibytes vs gigabytes). If they are then why would a scalar value of > 7236 not able to accomodate the pending tasks? > > -- > Andrew Jorgensen > @ajorgensen >
