Hi Andrew - this version added a fix for a stale weight issue that can occur on cgroup migration, and some fixups recommended by Ying Huang. There is an additional patch about the stale weight due to the introduction of atomics that may warrant some explicit scrutiny before pulling. v3: stale value fix, bulk allocator fixups, sysfs simplification --- 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 (High level results, 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. We implement sysfs entries for "system global" weights which can be set by a daemon or administrator. There are 3 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 information to be used during boot to set reasonable system default values based on the memory configuration of the system discovered at boot or during device 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 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 v3: - MAJOR: Changes cur_weight to be an atomic and to carry the current interleave node+weight, rather than just weight this fixes a stale-weight issue when a rebind occurs. - minor doc updates - sysfs: remove module_exit path, not needed - sysfs: allocate node_attrs rather than static MAX_NUMNODES array - interleave_nodes: handle cur_weight=0 conditions explicitly - bulk allocator: prev_node should be initialized to me->il_prev - bulk alloactor: weight collection logic style fixes - bulk allocator: bulk allocator for loop style fixes - bulk allocator: corner case fixes for resume_node/weight v2: - MAJOR: Torture tested bulk allocator, fixed edge conditions tracking the next me->il_node. Added documentation. Prior version was stable, but the resulting me->il_node could be wrong under certain circumstances. - naming: iw_table_mtx -> iw_table_lock - RCU: use synchronize+kfree and simplify the weight structure - default: remove default table, since it's static for now - sysfs setup: simplify setup, if table==NULL presume 1's - node_store: only allocate (sizeof(u8) * nr_node_ids) - allocators: update to deal with NULL table pointer - read_once: __builtin_memcpy -> memcpy - formatting v1: - RCU: This version protects the weight array with RCU. - ktest fix: proper include (types.h) in uapi header - doc: make mpol_params in docs reflect definition - doc: mempolicy.c comments in MPOL_WEIGHTED_INTERLEAVE patch - Dropped task-local weights and syscalls from the proposal until affirmative use cases for task-local weights appear. Link: https://lore.kernel.org/linux-mm/20240103224209.2541-1-gregory.price@xxxxxxxxxxxx/ ===================================================================== Performance tests - MLC >From - Ravi Jonnalagadda <ravis.opensrc@xxxxxxxxxx> 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 <gregory.price@xxxxxxxxxxxx> 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 <hyeongtak.ji@xxxxxx> 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 behavior like standard 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 <gregory.price@xxxxxxxxxxxx> Suggested-by: Johannes Weiner <hannes@xxxxxxxxxxx> Suggested-by: Hasan Al Maruf <hasanalmaruf@xxxxxx> Suggested-by: Hao Wang <haowang3@xxxxxx> Suggested-by: Ying Huang <ying.huang@xxxxxxxxx> Suggested-by: Dan Williams <dan.j.williams@xxxxxxxxx> Suggested-by: Michal Hocko <mhocko@xxxxxxxx> Suggested-by: Zhongkun He <hezhongkun.hzk@xxxxxxxxxxxxx> Suggested-by: Frank van der Linden <fvdl@xxxxxxxxxx> Suggested-by: John Groves <john@xxxxxxxxxxxxxx> Suggested-by: Vinicius Tavares Petrucci <vtavarespetr@xxxxxxxxxx> Suggested-by: Srinivasulu Thanneeru <sthanneeru@xxxxxxxxxx> Suggested-by: Ravi Jonnalagadda <ravis.opensrc@xxxxxxxxxx> Suggested-by: Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx> Suggested-by: Hyeongtak Ji <hyeongtak.ji@xxxxxx> Suggested-by: Andi Kleen <ak@xxxxxxxxxxxxxxx> Signed-off-by: Gregory Price <gregory.price@xxxxxxxxxxxx> 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: change cur_il_weight to atomic and carry the node with it 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/mempolicy.h | 3 + include/uapi/linux/mempolicy.h | 1 + mm/mempolicy.c | 551 +++++++++++++++++- 6 files changed, 579 insertions(+), 14 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
