lkml.org 
[lkml]   [2012]   [Nov]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC PATCH 0/8][Sorted-buddy] mm: Linux VM Infrastructure to support Memory Power Management
On Wed, Nov 07, 2012 at 01:22:13AM +0530, Srivatsa S. Bhat wrote:
> ------------------------------------------------------------
>
> Today memory subsystems are offer a wide range of capabilities for managing
> memory power consumption. As a quick example, if a block of memory is not
> referenced for a threshold amount of time, the memory controller can decide to
> put that chunk into a low-power content-preserving state. And the next
> reference to that memory chunk would bring it back to full power for read/write.
> With this capability in place, it becomes important for the OS to understand
> the boundaries of such power-manageable chunks of memory and to ensure that
> references are consolidated to a minimum number of such memory power management
> domains.
>

How much power is saved?

> ACPI 5.0 has introduced MPST tables (Memory Power State Tables) [5] so that
> the firmware can expose information regarding the boundaries of such memory
> power management domains to the OS in a standard way.
>

I'm not familiar with the ACPI spec but is there support for parsing of
MPST and interpreting the associated ACPI events? For example, if ACPI
fires an event indicating that a memory power node is to enter a low
state then presumably the OS should actively migrate pages away -- even
if it's going into a state where the contents are still refreshed
as exiting that state could take a long time.

I did not look closely at the patchset at all because it looked like the
actual support to use it and measure the benefit is missing.

> How can Linux VM help memory power savings?
>
> o Consolidate memory allocations and/or references such that they are
> not spread across the entire memory address space. Basically area of memory
> that is not being referenced, can reside in low power state.
>

Which the series does not appear to do.

> o Support targeted memory reclaim, where certain areas of memory that can be
> easily freed can be offlined, allowing those areas of memory to be put into
> lower power states.
>

Which the series does not appear to do judging from this;

include/linux/mm.h | 38 +++++++
include/linux/mmzone.h | 52 +++++++++
mm/compaction.c | 8 +
mm/page_alloc.c | 263 ++++++++++++++++++++++++++++++++++++++++++++----
mm/vmstat.c | 59 ++++++++++-

This does not appear to be doing anything with reclaim and not enough with
compaction to indicate that the series actively manages memory placement
in response to ACPI events.

Further in section 5.2.21.4 the spec says that power node regions can
overlap (but are not hierarchal for some reason) but have no gaps yet the
structure you use to represent is assumes there can be gaps and there are
no overlaps. Again, this is just glancing at the spec and a quick skim of
the patches so maybe I missed something that explains why this structure
is suitable.

It seems to me that superficially the VM implementation for the support
would have

a) Involved a tree that managed the overlapping regions (even if it's
not hierarchal it feels more sensible) and picked the highest-power-state
common denominator in the tree. This would only be allocated if support
for MPST is available.
b) Leave memory allocations and reclaim as they are in the active state.
c) Use a "sticky" migrate list MIGRATE_LOWPOWER for regions that are in lower
power but still usable with a latency penalty. This might be a single
migrate type but could also be a parallel set of free_area called
free_area_lowpower that is only used when free_area is depleted and in
the very slow path of the allocator.
d) Use memory hot-remove for power states where the refresh rates were
not constant

and only did anything expensive in response to an ACPI event -- none of
the fast paths should be touched.

When transitioning to the low power state, memory should be migrated in
a vaguely similar fashion to what CMA does. For low-power, migration
failure is acceptable. If contents are not preserved, ACPI needs to know
if the migration failed because it cannot enter that power state.

For any of this to be worthwhile, low power states would need to be achieved
for long periods of time because that migration is not free.

> Memory Regions:
> ---------------
>
> "Memory Regions" is a way of capturing the boundaries of power-managable
> chunks of memory, within the MM subsystem.
>
> Short description of the "Sorted-buddy" design:
> -----------------------------------------------
>
> In this design, the memory region boundaries are captured in a parallel
> data-structure instead of fitting regions between nodes and zones in the
> hierarchy. Further, the buddy allocator is altered, such that we maintain the
> zones' freelists in region-sorted-order and thus do page allocation in the
> order of increasing memory regions.

Implying that this sorting has to happen in the either the alloc or free
fast path.

> (The freelists need not be fully
> address-sorted, they just need to be region-sorted. Patch 6 explains this
> in more detail).
>
> The idea is to do page allocation in increasing order of memory regions
> (within a zone) and perform page reclaim in the reverse order, as illustrated
> below.
>
> ---------------------------- Increasing region number---------------------->
>
> Direction of allocation---> <---Direction of reclaim
>

Compaction will work against this because it uses a PFN walker to isolate
free pages and will ignore memory regions. If pageblocks were used, it
could take that into account at least.

> The sorting logic (to maintain freelist pageblocks in region-sorted-order)
> lies in the page-free path and not the page-allocation path and hence the
> critical page allocation paths remain fast.

Page free can be a critical path for application performance as well.
Think network buffer heavy alloc and freeing of buffers.

However, migratetype information is already looked up for THP so ideally
power awareness would piggyback on it.

> Moreover, the heart of the page
> allocation algorithm itself remains largely unchanged, and the region-related
> data-structures are optimized to avoid unnecessary updates during the
> page-allocator's runtime.
>
> Advantages of this design:
> --------------------------
> 1. No zone-fragmentation (IOW, we don't create more zones than necessary) and
> hence we avoid its associated problems (like too many zones, extra page
> reclaim threads, question of choosing watermarks etc).
> [This is an advantage over the "Hierarchy" design]
>
> 2. Performance overhead is expected to be low: Since we retain the simplicity
> of the algorithm in the page allocation path, page allocation can
> potentially remain as fast as it would be without memory regions. The
> overhead is pushed to the page-freeing paths which are not that critical.
>
>
> Results:
> =======
>
> Test setup:
> -----------
> This patchset applies cleanly on top of 3.7-rc3.
>
> x86 dual-socket quad core HT-enabled machine booted with mem=8G
> Memory region size = 512 MB
>
> Functional testing:
> -------------------
>
> Ran pagetest, a simple C program that allocates and touches a required number
> of pages.
>
> Below is the statistics from the regions within ZONE_NORMAL, at various sizes
> of allocations from pagetest.
>
> Present pages | Free pages at various allocations |
> | start | 512 MB | 1024 MB | 2048 MB |
> Region 0 16 | 0 | 0 | 0 | 0 |
> Region 1 131072 | 87219 | 8066 | 7892 | 7387 |
> Region 2 131072 | 131072 | 79036 | 0 | 0 |
> Region 3 131072 | 131072 | 131072 | 79061 | 0 |
> Region 4 131072 | 131072 | 131072 | 131072 | 0 |
> Region 5 131072 | 131072 | 131072 | 131072 | 79051 |
> Region 6 131072 | 131072 | 131072 | 131072 | 131072 |
> Region 7 131072 | 131072 | 131072 | 131072 | 131072 |
> Region 8 131056 | 105475 | 105472 | 105472 | 105472 |
>
> This shows that page allocation occurs in the order of increasing region
> numbers, as intended in this design.
>
> Performance impact:
> -------------------
>
> Kernbench results didn't show much of a difference between the performance
> of vanilla 3.7-rc3 and this patchset.
>
>
> Todos:
> =====
>
> 1. Memory-region aware page-reclamation:
> ----------------------------------------
>
> We would like to do page reclaim in the reverse order of page allocation
> within a zone, ie., in the order of decreasing region numbers.
> To achieve that, while scanning lru pages to reclaim, we could potentially
> look for pages belonging to higher regions (considering region boundaries)
> or perhaps simply prefer pages of higher pfns (and skip lower pfns) as
> reclaim candidates.
>

This would disrupting LRU ordering and if those pages were recently
allocated and you force a situation where swap has to be used then any
saving in low memory will be lost by having to access the disk instead.

> 2. Compile-time exclusion of Memory Power Management, and extending the
> support to also work with other features such as Mem cgroups, kexec etc.
>

Compile-time exclusion is pointless because it'll be always activated by
distribution configs. Support for MPST should be detected at runtime and

3. ACPI support to actually use this thing and validate the design is
compatible with the spec and actually works in hardware

--
Mel Gorman
SUSE Labs


\
 
 \ /
  Last update: 2012-11-08 19:41    [W:0.148 / U:2.972 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site