Re: [PATCH v5 0/5] mm: Randomize free memory

From: Dan Williams
Date: Mon Dec 17 2018 - 11:32:27 EST


On Mon, Dec 17, 2018 at 2:12 AM Rafael J. Wysocki <rjw@xxxxxxxxxxxxx> wrote:
>
> On Saturday, December 15, 2018 2:48:30 AM CET Dan Williams wrote:
> > Changes since v4: [1]
> > * Default the randomization to off and enable it dynamically based on
> > the detection of a memory side cache advertised by platform firmware.
> > In the case of x86 this enumeration comes from the ACPI HMAT. (Michal
> > and Mel)
> > * Improve the changelog of the patch that introduces the shuffling to
> > clarify the motivation and better explain the tradeoffs. (Michal and
> > Mel)
> > * Include the required HMAT enabling in the series.
> >
> > [1]: https://lkml.kernel.org/r/153922180166.838512.8260339805733812034.stgit@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
> >
> > ---
> >
> > Quote patch 3:
> >
> > Randomization of the page allocator improves the average utilization of
> > a direct-mapped memory-side-cache. Memory side caching is a platform
> > capability that Linux has been previously exposed to in HPC
> > (high-performance computing) environments on specialty platforms. In
> > that instance it was a smaller pool of high-bandwidth-memory relative to
> > higher-capacity / lower-bandwidth DRAM. Now, this capability is going to
> > be found on general purpose server platforms where DRAM is a cache in
> > front of higher latency persistent memory [2].
> >
> > Robert offered an explanation of the state of the art of Linux
> > interactions with memory-side-caches [3], and I copy it here:
> >
> > It's been a problem in the HPC space:
> > http://www.nersc.gov/research-and-development/knl-cache-mode-performance-coe/
> >
> > A kernel module called zonesort is available to try to help:
> > https://software.intel.com/en-us/articles/xeon-phi-software
> >
> > and this abandoned patch series proposed that for the kernel:
> > https://lkml.org/lkml/2017/8/23/195
> >
> > Dan's patch series doesn't attempt to ensure buffers won't conflict, but
> > also reduces the chance that the buffers will. This will make performance
> > more consistent, albeit slower than "optimal" (which is near impossible
> > to attain in a general-purpose kernel). That's better than forcing
> > users to deploy remedies like:
> > "To eliminate this gradual degradation, we have added a Stream
> > measurement to the Node Health Check that follows each job;
> > nodes are rebooted whenever their measured memory bandwidth
> > falls below 300 GB/s."
> >
> > A replacement for zonesort was merged upstream in commit cc9aec03e58f
> > "x86/numa_emulation: Introduce uniform split capability". With this
> > numa_emulation capability, memory can be split into cache sized
> > ("near-memory" sized) numa nodes. A bind operation to such a node, and
> > disabling workloads on other nodes, enables full cache performance.
> > However, once the workload exceeds the cache size then cache conflicts
> > are unavoidable. While HPC environments might be able to tolerate
> > time-scheduling of cache sized workloads, for general purpose server
> > platforms, the oversubscribed cache case will be the common case.
> >
> > The worst case scenario is that a server system owner benchmarks a
> > workload at boot with an un-contended cache only to see that performance
> > degrade over time, even below the average cache performance due to
> > excessive conflicts. Randomization clips the peaks and fills in the
> > valleys of cache utilization to yield steady average performance.
> >
> > See patch 3 for more details.
> >
> > [2]: https://itpeernetwork.intel.com/intel-optane-dc-persistent-memory-operating-modes/
> > [3]: https://lkml.org/lkml/2018/9/22/54
>
> Has this hibernation been tested with this series applied?

It has not. Is QEMU sufficient? What's your concern?