(v5: style, retry interleave w/ mems_allowed cookie
fix sparse warnings, style, review tags)
-----
Weighted interleave is a new interleave policy intended to make
use of heterogeneous memory environments appearing with CXL.
The existing interleave mechanism does an even round-robin
distribution of memory across all nodes in a nodemask, while
weighted interleave distributes memory across nodes according
to a provided weight. (Weight = # of page allocations per round)
Weighted interleave is intended to reduce average latency when
bandwidth is pressured - therefore increasing total throughput.
In other words: It allows greater use of the total available
bandwidth in a heterogeneous hardware environment (different
hardware provides different bandwidth capacity).
As bandwidth is pressured, latency increases - first linearly
and then exponentially. By keeping bandwidth usage distributed
according to available bandwidth, we therefore can reduce the
average latency of a cacheline fetch.
A good explanation of the bandwidth vs latency response curve:
https://mahmoudhatem.wordpress.com/2017/11/07/memory-bandwidth-vs-latency-response-curve/
From the article:
```
Constant region:
The latency response is fairly constant for the first 40%
of the sustained bandwidth.
Linear region:
In between 40% to 80% of the sustained bandwidth, the
latency response increases almost linearly with the bandwidth
demand of the system due to contention overhead by numerous
memory requests.
Exponential region:
Between 80% to 100% of the sustained bandwidth, the memory
latency is dominated by the contention latency which can be
as much as twice the idle latency or more.
Maximum sustained bandwidth :
Is 65% to 75% of the theoretical maximum bandwidth.
```
As a general rule of thumb:
* If bandwidth usage is low, latency does not increase. It is
optimal to place data in the nearest (lowest latency) device.
* If bandwidth usage is high, latency increases. It is optimal
to place data such that bandwidth use is optimized per-device.
This is the top line goal: Provide a user a mechanism to target using
the "maximum sustained bandwidth" of each hardware component in a
heterogenous memory system.
For example, the stream benchmark demonstrates that 1:1 (default)
interleave is actively harmful, while weighted interleave can be
beneficial. Default interleave distributes data such that too much
pressure is placed on devices with lower available bandwidth.
Stream Benchmark (vs DRAM, 1 Socket + 1 CXL Device)
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
Targeted weights : +2.5% to +4% (consistently better than DRAM)
Global means the task-policy was set (set_mempolicy), while targeted
means VMA policies were set (mbind2). We see weighted interleave
is not always beneficial when applied globally, but is always
beneficial when applied to bandwidth-driving memory regions.
There are 4 patches in this set:
1) Implement system-global interleave weights as sysfs extension
in mm/mempolicy.c. These weights are RCU protected, and a
default weight set is provided (all weights are 1 by default).
In future work, we intend to expose an interface for HMAT/CDAT
code to set reasonable default values based on the memory
configuration of the system discovered at boot/hotplug.
2) A mild refactor of some interleave-logic for re-use in the
new weighted interleave logic.
3) MPOL_WEIGHTED_INTERLEAVE extension for set_mempolicy/mbind
4) Protect interleave logic (weighted and normal) with the
mems_allowed seq cookie. If the nodemask changes while
accessing it during a rebind, just retry the access.
Included below are some performance and LTP test information,
and a sample numactl branch which can be used for testing.
= Performance summary =
(tests may have different configurations, see extended info below)
1) MLC (W2) : +38% over DRAM. +264% over default interleave.
MLC (W5) : +40% over DRAM. +226% over default interleave.
2) Stream : -6% to +4% over DRAM, +430% over default interleave.
3) XSBench : +19% over DRAM. +47% over default interleave.
= LTP Testing Summary =
existing mempolicy & mbind tests: pass
mempolicy & mbind + weighted interleave (global weights): pass
= version history
v5:
- style fixes
- mems_allowed cookie protection to detect rebind issues,
prevents spurious allocation failures and/or mis-allocations
- sparse warning fixes related to __rcu on local variables
=====================================================================
Performance tests - MLC
From - Ravi Jonnalagadda <[email protected]>
Hardware: Single-socket, multiple CXL memory expanders.
Workload: W2
Data Signature: 2:1 read:write
DRAM only bandwidth (GBps): 298.8
DRAM + CXL (default interleave) (GBps): 113.04
DRAM + CXL (weighted interleave)(GBps): 412.5
Gain over DRAM only: 1.38x
Gain over default interleave: 2.64x
Workload: W5
Data Signature: 1:1 read:write
DRAM only bandwidth (GBps): 273.2
DRAM + CXL (default interleave) (GBps): 117.23
DRAM + CXL (weighted interleave)(GBps): 382.7
Gain over DRAM only: 1.4x
Gain over default interleave: 2.26x
=====================================================================
Performance test - Stream
From - Gregory Price <[email protected]>
Hardware: Single socket, single CXL expander
numactl extension: https://github.com/gmprice/numactl/tree/weighted_interleave_master
Summary: 64 threads, ~18GB workload, 3GB per array, executed 100 times
Default interleave : -78% (slower than DRAM)
Global weighting : -6% to +4% (workload dependant)
mbind2 weights : +2.5% to +4% (consistently better than DRAM)
dram only:
numactl --cpunodebind=1 --membind=1 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Function Direction BestRateMBs AvgTime MinTime MaxTime
Copy: 0->0 200923.2 0.032662 0.031853 0.033301
Scale: 0->0 202123.0 0.032526 0.031664 0.032970
Add: 0->0 208873.2 0.047322 0.045961 0.047884
Triad: 0->0 208523.8 0.047262 0.046038 0.048414
CXL-only:
numactl --cpunodebind=1 -w --membind=2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 22209.7 0.288661 0.288162 0.289342
Scale: 0->0 22288.2 0.287549 0.287147 0.288291
Add: 0->0 24419.1 0.393372 0.393135 0.393735
Triad: 0->0 24484.6 0.392337 0.392083 0.394331
Based on the above, the optimal weights are ~9:1
echo 9 > /sys/kernel/mm/mempolicy/weighted_interleave/node1
echo 1 > /sys/kernel/mm/mempolicy/weighted_interleave/node2
default interleave:
numactl --cpunodebind=1 --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 44666.2 0.143671 0.143285 0.144174
Scale: 0->0 44781.6 0.143256 0.142916 0.143713
Add: 0->0 48600.7 0.197719 0.197528 0.197858
Triad: 0->0 48727.5 0.197204 0.197014 0.197439
global weighted interleave:
numactl --cpunodebind=1 -w --interleave=1,2 ./stream_c.exe --ntimes 100 --array-size 400M --malloc
Copy: 0->0 190085.9 0.034289 0.033669 0.034645
Scale: 0->0 207677.4 0.031909 0.030817 0.033061
Add: 0->0 202036.8 0.048737 0.047516 0.053409
Triad: 0->0 217671.5 0.045819 0.044103 0.046755
targted regions w/ global weights (modified stream to mbind2 malloc'd regions))
numactl --cpunodebind=1 --membind=1 ./stream_c.exe -b --ntimes 100 --array-size 400M --malloc
Copy: 0->0 205827.0 0.031445 0.031094 0.031984
Scale: 0->0 208171.8 0.031320 0.030744 0.032505
Add: 0->0 217352.0 0.045087 0.044168 0.046515
Triad: 0->0 216884.8 0.045062 0.044263 0.046982
=====================================================================
Performance tests - XSBench
From - Hyeongtak Ji <[email protected]>
Hardware: Single socket, Single CXL memory Expander
NUMA node 0: 56 logical cores, 128 GB memory
NUMA node 2: 96 GB CXL memory
Threads: 56
Lookups: 170,000,000
Summary: +19% over DRAM. +47% over default interleave.
Performance tests - XSBench
1. dram only
$ numactl -m 0 ./XSBench -s XL –p 5000000
Runtime: 36.235 seconds
Lookups/s: 4,691,618
2. default interleave
$ numactl –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 55.243 seconds
Lookups/s: 3,077,293
3. weighted interleave
numactl –w –i 0,2 ./XSBench –s XL –p 5000000
Runtime: 29.262 seconds
Lookups/s: 5,809,513
=====================================================================
LTP Tests: https://github.com/gmprice/ltp/tree/mempolicy2
= Existing tests
set_mempolicy, get_mempolicy, mbind
MPOL_WEIGHTED_INTERLEAVE added manually to test basic functionality
but did not adjust tests for weighting. Basically the weights were
set to 1, which is the default, and it should behave the same as
MPOL_INTERLEAVE if logic is correct.
== set_mempolicy01 : passed 18, failed 0
== set_mempolicy02 : passed 10, failed 0
== set_mempolicy03 : passed 64, failed 0
== set_mempolicy04 : passed 32, failed 0
== set_mempolicy05 - n/a on non-x86
== set_mempolicy06 : passed 10, failed 0
this is set_mempolicy02 + MPOL_WEIGHTED_INTERLEAVE
== set_mempolicy07 : passed 32, failed 0
set_mempolicy04 + MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy01 : passed 12, failed 0
change: added MPOL_WEIGHTED_INTERLEAVE
== get_mempolicy02 : passed 2, failed 0
== mbind01 : passed 15, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind02 : passed 4, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind03 : passed 16, failed 0
added MPOL_WEIGHTED_INTERLEAVE
== mbind04 : passed 48, failed 0
added MPOL_WEIGHTED_INTERLEAVE
=====================================================================
numactl (set_mempolicy) w/ global weighting test
numactl fork: https://github.com/gmprice/numactl/tree/weighted_interleave_master
command: numactl -w --interleave=0,1 ./eatmem
result (weights 1:1):
0176a000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=32897 N1=32896 kernelpagesize_kB=4
7fceeb9ff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=32768 N1=32769 kernelpagesize_kB=4
50% distribution is correct
result (weights 5:1):
01b14000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=54828 N1=10965 kernelpagesize_kB=4
7f47a1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=54614 N1=10923 kernelpagesize_kB=4
16.666% distribution is correct
result (weights 1:5):
01f07000 weighted interleave:0-1 heap anon=65793 dirty=65793 active=0 N0=10966 N1=54827 kernelpagesize_kB=4
7f17b1dff000 weighted interleave:0-1 anon=65537 dirty=65537 active=0 N0=10923 N1=54614 kernelpagesize_kB=4
16.666% distribution is correct
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void)
{
char* mem = malloc(1024*1024*256);
memset(mem, 1, 1024*1024*256);
for (int i = 0; i < ((1024*1024*256)/4096); i++)
{
mem = malloc(4096);
mem[0] = 1;
}
printf("done\n");
getchar();
return 0;
}
=====================================================================
Suggested-by: Gregory Price <[email protected]>
Suggested-by: Johannes Weiner <[email protected]>
Suggested-by: Hasan Al Maruf <[email protected]>
Suggested-by: Hao Wang <[email protected]>
Suggested-by: Ying Huang <[email protected]>
Suggested-by: Dan Williams <[email protected]>
Suggested-by: Michal Hocko <[email protected]>
Suggested-by: Zhongkun He <[email protected]>
Suggested-by: Frank van der Linden <[email protected]>
Suggested-by: John Groves <[email protected]>
Suggested-by: Vinicius Tavares Petrucci <[email protected]>
Suggested-by: Srinivasulu Thanneeru <[email protected]>
Suggested-by: Ravi Jonnalagadda <[email protected]>
Suggested-by: Jonathan Cameron <[email protected]>
Suggested-by: Hyeongtak Ji <[email protected]>
Suggested-by: Andi Kleen <[email protected]>
Signed-off-by: Gregory Price <[email protected]>
Gregory Price (3):
mm/mempolicy: refactor a read-once mechanism into a function for
re-use
mm/mempolicy: introduce MPOL_WEIGHTED_INTERLEAVE for weighted
interleaving
mm/mempolicy: protect task interleave functions with
tsk->mems_allowed_seq
Rakie Kim (1):
mm/mempolicy: implement the sysfs-based weighted_interleave interface
.../ABI/testing/sysfs-kernel-mm-mempolicy | 4 +
...fs-kernel-mm-mempolicy-weighted-interleave | 25 +
.../admin-guide/mm/numa_memory_policy.rst | 9 +
include/linux/sched.h | 1 +
include/uapi/linux/mempolicy.h | 1 +
mm/mempolicy.c | 488 +++++++++++++++++-
6 files changed, 513 insertions(+), 15 deletions(-)
create mode 100644 Documentation/ABI/testing/sysfs-kernel-mm-mempolicy
create mode 100644 Documentation/ABI/testing/sysfs-kernel-mm-mempolicy-weighted-interleave
--
2.39.1
move the use of barrier() to force policy->nodemask onto the stack into
a function `read_once_policy_nodemask` so that it may be re-used.
Suggested-by: "Huang, Ying" <[email protected]>
Signed-off-by: Gregory Price <[email protected]>
Reviewed-by: "Huang, Ying" <[email protected]>
---
mm/mempolicy.c | 26 ++++++++++++++++----------
1 file changed, 16 insertions(+), 10 deletions(-)
diff --git a/mm/mempolicy.c b/mm/mempolicy.c
index 41e58c4c0d01..697f2a791c24 100644
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -1909,6 +1909,20 @@ unsigned int mempolicy_slab_node(void)
}
}
+static unsigned int read_once_policy_nodemask(struct mempolicy *pol,
+ nodemask_t *mask)
+{
+ /*
+ * barrier stabilizes the nodemask locally so that it can be iterated
+ * over safely without concern for changes. Allocators validate node
+ * selection does not violate mems_allowed, so this is safe.
+ */
+ barrier();
+ memcpy(mask, &pol->nodes, sizeof(nodemask_t));
+ barrier();
+ return nodes_weight(*mask);
+}
+
/*
* Do static interleaving for interleave index @ilx. Returns the ilx'th
* node in pol->nodes (starting from ilx=0), wrapping around if ilx
@@ -1916,20 +1930,12 @@ unsigned int mempolicy_slab_node(void)
*/
static unsigned int interleave_nid(struct mempolicy *pol, pgoff_t ilx)
{
- nodemask_t nodemask = pol->nodes;
+ nodemask_t nodemask;
unsigned int target, nnodes;
int i;
int nid;
- /*
- * The barrier will stabilize the nodemask in a register or on
- * the stack so that it will stop changing under the code.
- *
- * Between first_node() and next_node(), pol->nodes could be changed
- * by other threads. So we put pol->nodes in a local stack.
- */
- barrier();
- nnodes = nodes_weight(nodemask);
+ nnodes = read_once_policy_nodemask(pol, &nodemask);
if (!nnodes)
return numa_node_id();
target = ilx % nnodes;
--
2.39.1