Re: [PATCH v3 0/9] mm: introduce Designated Movable Blocks

From: Doug Berger
Date: Wed Dec 14 2022 - 19:19:12 EST

On 11/18/2022 9:05 AM, Mel Gorman wrote:
On Wed, Nov 02, 2022 at 03:33:53PM -0700, Doug Berger wrote:

Hotplugs requirements are somewhat different, the primary motivation that
I'm aware of is being able to guarantee they can be offlined, particularly
nodes, which can be done in some circumstances. Generally hotplug does
not care what uses the memory as long as it can be removed later. The
requirements for restricted access to high speed memory is different.

This is effectively the same requirement that an implementation of
'reusable' reserved memory has. A driver that owns reserved memory may not
care what uses the memory as long as the memory can be reclaimed when the
driver needs it. This is functionally analogous to memory hotplug. Reserved
memory that is 'reusable' and compatible with 'shared-dma-pool' uses the CMA
implementation, but there is room for an alternative implementation that
shares the memory more aggressively. This is a separate motivator for
Designated Movable Block support, but I am deferring that discussion since
it is desirable to have a more extended debate over APIs and such.


There needs to be a better explanation as to why CMA cannot be used or more
importantly why page_alloc.shuffle= should not be used (more on that later.
It's not clear how a movable zone shares memory more aggressively than CMA
would. Both have the problem that the if protected range is too large that
premature memory exhaustion can occur for kernel allocations.
The pages within a CMA pool can be allocated by either the CMA allocator (i.e. alloc_contig_range()) for use by the kernel (i.e. drivers) or by the page allocator to satisfy movable requests (i.e. predominantly user space). However, the page allocator is constrained by rules that make allocations from MIGRATE_CMA free lists a secondary consideration when the memory is available elsewhere. This tends toward keeping pages in a CMA pool free while user processes consume more memory outside of the CMA pool that could have been used more generally by the kernel.
Pages on the MIGRATE_MOVABLE free list in ZONE_MOVABLE (i.e. all of the pages in the movable zone) are the first choice for satisfying movable requests. This allows user space to make full use of a 'reusable' reserved memory range that isn't actively used by a driver, which is what I mean by "more aggressive". When the driver wants to reclaim its 'reusable' reserved memory range it can use alloc_contig_range() to force any movable allocations out of the range perhaps into memory that could have been used more generally by the kernel. Such a reclamation may be more time consuming to complete since the pages are more likely to be in use than if they were on the MIGRATE_CMA free list, but no pages go unused by either the driver or user space.

I was not familiar with page_alloc.shuffle, but it may very well have a role to play here.



There is a high degree of uncertainity of how these regions are to be
used by applications to get access to the high speed memory, to quote

I'm not certain what is typical because these systems are highly
configurable and Broadcom's customers have different ideas about
application processing.

...

The Designated Movable Block concept introduced here has the
potential to offer useful services to different constituencies. I
tried to highlight this in my V1 patch set with the hope of
attracting some interest, but it can complicate the overall
discussion, so I would like to maybe narrow the discussion here. It
may be good to keep them in mind when assessing the overall value,
but perhaps the "other opportunities" can be covered as a follow
on discussion.

I note the "potential" part here because we don't actually know.

I used "potential" here not as in "it might be useful", but rather that
"different constituencies (i.e. people outside of the Broadcom ecosystem)
might also find them useful".


That's very vague unfortunately.
As an example, there have been submissions related to memory hotplug that could have used this capability if it existed at the time. In fact the comments for [1] incorrectly assumed the movablecore= behavior proposed here was already implemented. Eventually those patch sets morphed into the movable_node implementation and modification of movablecore= was dropped.

The point I intended to make is that in addition to the use case identified here (i.e. improved bandwidth utilization on multiple memory controller systems) Designated Movable Blocks as a general mechanism could be useful to people I don't know to solve problems of which I am not aware (e.g. memory hot unplugging, reusable reserved memory, ???). I offer as evidence that I am not the only person to conceive of the concept.

[1] https://lore.kernel.org/all/1374220774-29974-21-git-send-email-tangchen@xxxxxxxxxxxxxx/


A
major limitation of ZONE_MOVABLE is that there is no way of controlling
access from userspace to restrict the high-speed memory to a designated
application, only to all applications in general. The primary interface
to control access to memory with different characteristics is mempolicies
which is NUMA orientated, not zone orientated. So, if there is a special
application that requires exclusive access, it's very difficult to configure
based on zones. Furthermore, page table pages mapping data located in the
high-speed region are stored in the slower memory which potentially impacts
the performance if the working set of the application exceeds TLB reach.
Finally, while there is mention that Broadcom may have some special
interface to determine what applications can use the high-speed region,
it's hardware-specific as opposed to something that belongs in the core mm.

I agree that keeping the high-speed memory in a local node and using "sticky"
pageblocks or CMA has limitations of its own but in itself, that does not
justify using ZONE_MOVABLE in my opinion. The statement that ARM can have
multiple controllers with equal distance and bandwidth (if I'm reading it
correctly) but places them in different zones.... that's just a bit weird if
there are no other addressing limitations. It's not obvious why ARM would do
that, but it also does not matter because it shouldn't be a core mm concern.

There appears to be some confusion regarding my explanation of multiple
memory controllers on a device like the BCM7278. There is no inherent
performance difference between the two memory controllers and their attached
DRAM. They merely provide the opportunity to perform memory accesses in
parallel for different physical address ranges. The physical address ranges
were selected by the SoC designers for reasons only known to them, but I'm
sure they had no consideration of zones in their decision making. The
selection of zones remains an artifact of the design of Linux.


Ok, so the channels are equal but the channels are not interleaved in
hardware so basically you are trying to implement software-based memory
channel interleaving?
I suppose that could be a fair characterization of the objective, though the approach taken here is very much a "poor man's" approach that attempts to improve things without requiring the "heavy lifting" required for a more complete solution.

That said, the use of page_alloc.shuffle is interesting here since it could improve the likelyhood of interleaving.


Since the BCM7278 contains a 4-core SMP cluster and each core can have
multiple outstanding memory transactions the speed of DDR transactions can
create a bottleneck for the system. If each memory controller has an
effective bandwidth of X then, provided the system memory accesses can be
distributed across both memory controllers, the combined effective bandwidth
can be additive (X + X = 2X). Of course the actual result is highly
dependent on the dependent clause "provided the system memory accesses can
be distributed across both memory controllers". The accesses do not need to
be evenly distributed to gain a benefit. We just want to reduce any idle
time on each memory controller.

It was observed that the monotonic zone layout for a non-NUMA system (like
this one) creates a bias for kernel space to use lower physical memory
addresses and user space to use higher physical memory addresses. Broadcom
customers requested the ability to locate movablecore memory within the
physical address range of each memory controller and reported that it
improved their system performance. Unfortunately, I do not have access to
their data and I doubt they would allow me to share it if I did. I don't
believe this is really about trying to optimize the performance of a
specific application as much as trying to prevent overall system performance
degradation from underutilized memory bandwidth.


So if I'm reading this right, the intent behind using ZONE_MOVABLE at
fixed address ranges is so most (but not all) user pages end up using
one controller and all kernel pages and some user pages use the other
controller. If kernel and userspace accesses are split 50/50, then the
memory bandwidth usage will be split across channels. However, if the
ratio of kernel:user accesses is large then the bandwidth usage will still
be assymetric.

For example, if there are 10 times more accesses to kernel pages then user
pages, then one channel will receive most of the traffic. The reverse
is also true but to a lesser extent as user pages can use all zones and
kernel accesses use a subset. Depending on the access pattern, creating
separate zones may not help at all and in some cases, could make the problem
worse. The trap is that it might happen to work for a fixed appliance like
a TV with a predictable workload, it may not work in the general case.

Splitting based on the __GFP_MOVABLE does not guarantee that idle time on
a memory controller can be reduced as it relies on the access pattern.
You are not reading it quite right. We could accomplish the split you describe on a BCM7278 SoC with two memory controllers using the existing movablecore=50% kernel parameter. This would create a ZONE_MOVABLE on the high address memory controller and a ZONE_DMA on the low address memory controller.

What is of interest to Broadcom customers is to better distribute user space accesses across each memory controller to improve the bandwidth available to user space dominated work flows. With no ZONE_MOVABLE, the BCM7278 SoC with 1GB of memory on each memory controller will place the 1GB on the low address memory controller in ZONE_DMA and the 1GB on the high address memory controller in ZONE_NORMAL. With this layout movable allocation requests will only fallback to the ZONE_DMA (low memory controller) once the ZONE_NORMAL (high memory controller) is sufficiently depleted of free memory.

Adding ZONE_MOVABLE memory above ZONE_NORMAL with the current movablecore behavior does not improve this situation other than forcing more kernel allocations off of the high memory controller. User space allocations are even more likely to be on the high memory controller.

The Designated Movable Block mechanism allows ZONE_MOVABLE memory to be located on the low memory controller to make it easier for user space allocations to land on the low memory controller. If ZONE_MOVABLE is only placed on the low memory controller then user space allocations can land in ZONE_NORMAL on the high memory controller, but only through fallback after ZONE_MOVABLE is sufficiently depleted of free memory which is just the reverse of the existing situation. The Designated Movable Block mechanism allows ZONE_MOVABLE memory to be located on each memory controller so that user space allocations have equal access to each memory controller until the ZONE_MOVABLE memory is depleted and fallback to other zones occurs.

To my knowledge Broadcom customers that are currently using the Designated Movable Block mechanism are relying on the somewhat random starting and stopping of parallel user space processes to produce a more random distribution of ZONE_MOVABLE allocations across multiple memory controllers, but the page_alloc.shuffle mechanism seems like it would be a good addition to promote this randomness. Even better, it appears that page_alloc.shuffle is already enabled in the GKI configuration.

You are of course correct that the access patterns make all of the difference and it is almost certain that one memory controller or the other will be saturated at any given time, but the intent is to increase the opportunity to use more of the total bandwidth made available by the multiple memory controllers.


There are already examples of where memory is physically "local" to
the CPU but has different bandwidth or latency including High Bandwidth
(HBM), Sub-NUMA Clustering (SNC), PMEM as a memory-life device and some
AMD EPYC Chips, particularly the first generation where a sockets memory
controllers had different distances. With the broadcom controllers,
it sounds like a local memory controller but the bandwidth available
differs. It's functionally equivalent to HBM.

The bandwidth available does not differ, but if too few transactions target
one of the memory controllers, that controllers bandwidth is underutilized.


This is also a limitation of the patch series. Lets say the bulk of
accesses are to user pages allocated in ZONE_MOVABLE which correlates to
one memory channel then one channel gets saturated anyway.

It also gets more complicated if there are more controllers because the
only division possible is between MOVABLE/everything else. An odd number
of channels will be difficult to split meaningfully.
The patch series allows Designated Movable Blocks to occupy a portion of each memory controller while allowing the ZONE_MOVABLE zone to span all of the memory controllers. In this way user pages allocated from ZONE_MOVABLE may be distributed across all of the memory controllers. Use of page_alloc.shuffle should improve the randomness of this distribution. Memory outside of Designated Movable Blocks on each memory controller can be outside ZONE_MOVABLE (e.g. ZONE_DMA and ZONE_NORMAL) and managed accordingly. An odd number of channels need not affect this.


The fact that the memory access is physically local to the CPU socket is
irrelevant when the characteristics of that locality differs. NUMA stands
for Non-Uniform Memory Access and if bandwidth to different address ranges
differs, then the system is inherently NUMA even if that is inconvenient.

The bandwidth to different address ranges does not differ. A single threaded
application should see no performance difference regardless of where its
memory is located. However, if multiple application threads are executing in
parallel and the memory is divided between the memory controllers they will
be able to do more work per unit of time than if the memory is predominantly
located on one memory controller.


And if multiple application threads dominantly access user pages then
splitting the zone will not necessarily help, particularly if ZONE_MOVABLE
is not filled as the bulk of the accesses will still use one memory channel.

In the appliance case, it doesn't matter if the intent is that "all
application data should use high bandwidth memory where possible and
the application phase behaviour is predictable" and that may very well
work fine for the users of the Broadcom platforms with multiple memory
controllers. It does not work at all for the general where access must
be restricted to a subset of tasks in a general system that can only be
controlled with memory policies.

The high bandwidth memory should be representated as a NUMA node, optionally
to create that node as ZONE_MOVABLE and relying on the zonelists to select
the movable zone as the first preference.

This patch set is fundamentally about greater control over the placement of
movablecore memory. The current implementation of movablecore requires all
of the ZONE_MOVABLE memory to be located at the highest physical addresses
of the system when CONFIG_NUMA is not set. Allowing the specification of a
base address allows greater flexibility on systems where there are benefits.


Unfortunately, while greater control of the ranges used by ZONE_MOVABLE
will help in some cases, it will not help in others and may be misleading.

If memory accesses need to be interleaved in software then the free lists
need to be broken up into arenas within a zone by some boundary whether
that boundary is is fixed-length ranges, memory sections, memory channels
or specified explicitly on the kernel command line. Any allocation type
can use any arena with tasks moving to another arena based on contention,
pageblock type availability or interleaving round-robin explicitly.
Unfortunately, it's non-trivial to implement and a *lot* of heavy lifting.

A somewhat awful hack would be to reorder top-level MAX_ORDER-1 list at
boot time. By default that list is ordered

1, 2, 3 ...... n-2, n-1, n

If at boot time it was reordered to be

1, n, 2, n-1, 3, n-2 ......

This would interleave all the early allocations across memory channels in
the case where channels are based on large contiguous physical ranges of
memory. Applications starting early would then interleave between channels
but after a period of time, it would be pseudo-random and it's weak.

A similar, and probably better, option is to look at what page_alloc.shuffle=
does and randomly shuffle the free lists to randomly select between the
memory channels. I haven't looked at the implementation recently and forget
how it works exactly. Maybe it would benefit from being able to take ranges
that should be special cased for shuffling, particularly at boot time to
order it "1, n, 2, n-1" as described above or allowing SHUFFLE_ORDER to
be a lower value. Either way, shuffling would achieve similar goals of
spreading allocations between channels without assuming that the access
ratio of kernel:user is close to 1:1.


I decided to implement this very simple multi-threaded application as a testcase to experiment with the concepts discussed here:
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <pthread.h>

#define BUF_SIZE (0x4000000)
#define THREADS (4)
#define COPY_COUNT (30)

void *thread_function( void *ptr );

int main()
{
pthread_t thread[THREADS];
int i, iret[THREADS];

for(i = 0; i < THREADS; i++)
iret[i] = pthread_create( &thread[i], NULL, thread_function, (void*) NULL);

for(i = 0; i < THREADS; i++)
pthread_join( thread[i], NULL);

for(i = 0; i < THREADS; i++)
printf("Thread %d returns: %d\n", i, iret[i]);
exit(0);
}

void *thread_function( void *ptr )
{
char *s, *d;
int i;

s = malloc(BUF_SIZE);
if (!s)
return NULL;

d = malloc(BUF_SIZE);
if (!d) {
free(s);
return NULL;
}

for (i = 0; i < COPY_COUNT; i++) {
memcpy(d, s, BUF_SIZE);
}
free(s);
free(d);
}

It meaninglessly moves data from one large dynamically allocated buffer to another a number of times without trying to be clever. I experimented with a Broadcom BCM7278 system with 1GB on each memory controller (i.e. 2GB total memory). The buffers were made large to render data caching meaningless and to require several pages to be allocated to populate the buffer.

With V3 of this patch set applied to a 6.1-rc1 kernel I observed these results:
With no movablecore kernel parameter specified:
# time /tmp/thread_test
Thread 1 returns: 0
Thread 2 returns: 0
Thread 3 returns: 0
Thread 4 returns: 0

real 0m4.047s
user 0m14.183s
sys 0m1.215s

With this additional kernel parameter "movablecore=600M":
# time /tmp/thread_test
Thread 0 returns: 0
Thread 1 returns: 0
Thread 2 returns: 0
Thread 3 returns: 0

real 0m4.068s
user 0m14.402s
sys 0m1.117s

With this additional kernel parameter "movablecore=600M@0x50000000":
# time /tmp/thread_test
Thread 0 returns: 0
Thread 1 returns: 0
Thread 2 returns: 0
Thread 3 returns: 0

real 0m4.010s
user 0m13.979s
sys 0m1.070s

However, with these additional kernel parameters "movablecore=300M@0x60000000,300M@0x320000000 page_alloc.shuffle=1":
# time /tmp/thread_test
Thread 0 returns: 0
Thread 1 returns: 0
Thread 2 returns: 0
Thread 3 returns: 0

real 0m3.173s
user 0m11.175s
sys 0m1.067s

These numbers show an over 20% improvement in performance of the test application when distributing ZONE_MOVABLE across both memory controllers.

Happy Holidays!
Doug