Re: [RFC v2 0/5] surface heterogeneous memory performance information
From: Balbir Singh
Date: Fri Jul 07 2017 - 02:28:37 EST
On Thu, 2017-07-06 at 15:52 -0600, Ross Zwisler wrote:
> ==== Quick Summary ====
> Platforms in the very near future will have multiple types of memory
> attached to a single CPU. These disparate memory ranges will have some
> characteristics in common, such as CPU cache coherence, but they can have
> wide ranges of performance both in terms of latency and bandwidth.
> For example, consider a system that contains persistent memory, standard
> DDR memory and High Bandwidth Memory (HBM), all attached to the same CPU.
> There could potentially be an order of magnitude or more difference in
> performance between the slowest and fastest memory attached to that CPU.
> With the current Linux code NUMA nodes are CPU-centric, so all the memory
> attached to a given CPU will be lumped into the same NUMA node. This makes
> it very difficult for userspace applications to understand the performance
> of different memory ranges on a given CPU.
> We solve this issue by providing userspace with performance information on
> individual memory ranges. This performance information is exposed via
> # grep . mem_tgt2/* mem_tgt2/local_init/* 2>/dev/null
Could you please explain these charactersitics, are they in the patches
How to these numbers compare to normal system memory?
> This allows applications to easily find the memory that they want to use.
> We expect that the existing NUMA APIs will be enhanced to use this new
> information so that applications can continue to use them to select their
> desired memory.
> This series is built upon acpica-1705:
> And you can find a working tree here:
> ==== Lots of Details ====
> This patch set is only concerned with CPU-addressable memory types, not
> on-device memory like what we have with Jerome Glisse's HMM series:
> This patch set works by enabling the new Heterogeneous Memory Attribute
> Table (HMAT) table, newly defined in ACPI 6.2. One major conceptual change
> in ACPI 6.2 related to this work is that proximity domains no longer need
> to contain a processor. We can now have memory-only proximity domains,
> which means that we can now have memory-only Linux NUMA nodes.
> Here is an example configuration where we have a single processor, one
> range of regular memory and one range of HBM:
> +---------------+ +----------------+
> | Processor | | Memory |
> | prox domain 0 +---+ prox domain 1 |
> | NUMA node 1 | | NUMA node 2 |
> +-------+-------+ +----------------+
> | HBM |
> | prox domain 2 |
> | NUMA node 0 |
> This gives us one initiator (the processor) and two targets (the two memory
> ranges). Each of these three has its own ACPI proximity domain and
> associated Linux NUMA node. Note also that while there is a 1:1 mapping
> from each proximity domain to each NUMA node, the numbers don't necessarily
> match up. Additionally we can have extra NUMA nodes that don't map back to
> ACPI proximity domains.
Could you expand on proximity domains, are they the same as node distance
or is this ACPI terminology for something more?
> The above configuration could also have the processor and one of the two
> memory ranges sharing a proximity domain and NUMA node, but for the
> purposes of the HMAT the two memory ranges will always need to be
> The overall goal of this series and of the HMAT is to allow users to
> identify memory using its performance characteristics. This can broadly be
> done in one of two ways:
> Option 1: Provide the user with a way to map between proximity domains and
> NUMA nodes and a way to access the HMAT directly (probably via
> /sys/firmware/acpi/tables). Then, through possibly a library and a daemon,
> provide an API so that applications can either request information about
> memory ranges, or request memory allocations that meet a given set of
> performance characteristics.
> Option 2: Provide the user with HMAT performance data directly in sysfs,
> allowing applications to directly access it without the need for the
> library and daemon.
> The kernel work for option 1 is started by patches 1-3. These just surface
> the minimal amount of information in sysfs to allow userspace to map
> between proximity domains and NUMA nodes so that the raw data in the HMAT
> table can be understood.
> Patches 4 and 5 enable option 2, adding performance information from the
> HMAT to sysfs. The second option is complicated by the amount of HMAT data
> that could be present in very large systems, so in this series we only
> surface performance information for local (initiator,target) pairings. The
> changelog for patch 5 discusses this in detail.
> The naming collision between Jerome's "Heterogeneous Memory Management
> (HMM)" and this "Heterogeneous Memory (HMEM)" series is unfortunate, but I
> was trying to stick with the word "Heterogeneous" because of the naming of
> the ACPI 6.2 Heterogeneous Memory Attribute Table table. Suggestions for
> better naming are welcome.