NUMA is a real issue but a) not all apps get affected and b) not all the solutions out there are actually affected by it.
At some point, it would be great if we can integrate into mesos something like http://www.open-mpi.org/projects/hwloc/ so that we can reason/optimize about such HW issues in a reasonable and portable manner. On Wed, Feb 11, 2015 at 6:53 AM, Chester Kuo <[email protected]> wrote: > Thank you ,Alex & Adam. > > I will look more from (1), as we did before pinning task on local memory > had better performance than remote memory. > > Chester > 2015/2/11 下午6:17 於 "Alex Rukletsov" <[email protected]> 寫道： > > > NUMA optimizations can be controversial [1], but it would be nice if you > > help out adding the optional support for it. > > > > [1] > > > > > https://engineering.linkedin.com/performance/optimizing-linux-memory-management-low-latency-high-throughput-databases > > > > On Wed, Feb 11, 2015 at 10:07 AM, Adam Bordelon <[email protected]> > > wrote: > > > > > Don't think anybody's actively working on it. See MESOS-314 > > > <https://issues.apache.org/jira/browse/MESOS-314> for "cpuset" > > > > > > On Tue, Feb 10, 2015 at 11:39 PM, Chester Kuo <[email protected]> > > > wrote: > > > > > > > Hi dev, > > > > > > > > Just checking with current Mesos and i didn't see it utilize NUMA > > > > system of memory/CPU allocation on slave. > > > > Are there anyone planning on this or correct me if this is supported. > > > > > > > > > > > > There is a feature called cpuset implemented in the linux kernel, > > > > which is very useful for HPC application on NUMA system. > > > > > > > > > > > > > > > > Chester > > > > > > > > > >
