-
Honggyu Kim authored
Patch series "DAMON based tiered memory management for CXL memory", v6. Introduction ============ With the advent of CXL/PCIe attached DRAM, which will be called simply as CXL memory in this cover letter, some systems are becoming more heterogeneous having memory systems with different latency and bandwidth characteristics. They are usually handled as different NUMA nodes in separate memory tiers and CXL memory is used as slow tiers because of its protocol overhead compared to local DRAM. In this kind of systems, we need to be careful placing memory pages on proper NUMA nodes based on the memory access frequency. Otherwise, some frequently accessed pages might reside on slow tiers and it makes performance degradation unexpectedly. Moreover, the memory access patterns can be changed at runtime. To handle this problem, we need a way to monitor the memory access patterns and migrate pages based on their access temperature. The DAMON(Data Access MONitor) framework and its DAMOS(DAMON-based Operation Schemes) can be useful features for monitoring and migrating pages. DAMOS provides multiple actions based on DAMON monitoring results and it can be used for proactive reclaim, which means swapping cold pages out with DAMOS_PAGEOUT action, but it doesn't support migration actions such as demotion and promotion between tiered memory nodes. This series supports two new DAMOS actions; DAMOS_MIGRATE_HOT for promotion from slow tiers and DAMOS_MIGRATE_COLD for demotion from fast tiers. This prevents hot pages from being stuck on slow tiers, which makes performance degradation and cold pages can be proactively demoted to slow tiers so that the system can increase the chance to allocate more hot pages to fast tiers. The DAMON provides various tuning knobs but we found that the proactive demotion for cold pages is especially useful when the system is running out of memory on its fast tier nodes. Our evaluation result shows that it reduces the performance slowdown compared to the default memory policy from 11% to 3~5% when the system runs under high memory pressure on its fast tier DRAM nodes. DAMON configuration =================== The specific DAMON configuration doesn't have to be in the scope of this patch series, but some rough idea is better to be shared to explain the evaluation result. The DAMON provides many knobs for fine tuning but its configuration file is generated by HMSDK[3]. It includes gen_config.py script that generates a json file with the full config of DAMON knobs and it creates multiple kdamonds for each NUMA node when the DAMON is enabled so that it can run hot/cold based migration for tiered memory. Evaluation Workload =================== The performance evaluation is done with redis[4], which is a widely used in-memory database and the memory access patterns are generated via YCSB[5]. We have measured two different workloads with zipfian and latest distributions but their configs are slightly modified to make memory usage higher and execution time longer for better evaluation. The idea of evaluation using these migrate_{hot,cold} actions covers system-wide memory management rather than partitioning hot/cold pages of a single workload. The default memory allocation policy creates pages to the fast tier DRAM node first, then allocates newly created pages to the slow tier CXL node when the DRAM node has insufficient free space. Once the page allocation is done then those pages never move between NUMA nodes. It's not true when using numa balancing, but it is not the scope of this DAMON based tiered memory management support. If the working set of redis can be fit fully into the DRAM node, then the redis will access the fast DRAM only. Since the performance of DRAM only is faster than partially accessing CXL memory in slow tiers, this environment is not useful to evaluate this patch series. To make pages of redis be distributed across fast DRAM node and slow CXL node to evaluate our migrate_{hot,cold} actions, we pre-allocate some cold memory externally using mmap and memset before launching redis-server. We assumed that there are enough amount of cold memory in datacenters as TMO[6] and TPP[7] papers mentioned. The evaluation sequence is as follows. 1. Turn on DAMON with DAMOS_MIGRATE_COLD action for DRAM node and DAMOS_MIGRATE_HOT action for CXL node. It demotes cold pages on DRAM node and promotes hot pages on CXL node in a regular interval. 2. Allocate a huge block of cold memory by calling mmap and memset at the fast tier DRAM node, then make the process sleep to make the fast tier has insufficient space for redis-server. 3. Launch redis-server and load prebaked snapshot image, dump.rdb. The redis-server consumes 52GB of anon pages and 33GB of file pages, but due to the cold memory allocated at 2, it fails allocating the entire memory of redis-server on the fast tier DRAM node so it partially allocates the remaining on the slow tier CXL node. The ratio of DRAM:CXL depends on the size of the pre-allocated cold memory. 4. Run YCSB to make zipfian or latest distribution of memory accesses to redis-server, then measure its execution time when it's completed. 5. Repeat 4 over 50 times to measure the average execution time for each run. 6. Increase the cold memory size then repeat goes to 2. For each test at 4 took about a minute so repeating it 50 times almost took about 1 hour for each test with a specific cold memory from 440GB to 500GB in 10GB increments for each evaluation. So it took about more than 10 hours for both zipfian and latest workloads to get the entire evaluation results. Repeating the same test set multiple times doesn't show much difference so I think it might be enough to make the result reliable. Evaluation Results ================== All the result values are normalized to DRAM-only execution time because the workload cannot be faster than DRAM-only unless the workload hits the peak bandwidth but our redis test doesn't go beyond the bandwidth limit. So the DRAM-only execution time is the ideal result without affected by the gap between DRAM and CXL performance difference. The NUMA node environment is as follows. node0 - local DRAM, 512GB with a CPU socket (fast tier) node1 - disabled node2 - CXL DRAM, 96GB, no CPU attached (slow tier) The following is the result of generating zipfian distribution to redis-server and the numbers are averaged by 50 times of execution. 1. YCSB zipfian distribution read only workload memory pressure with cold memory on node0 with 512GB of local DRAM. ====================+================================================+========= | cold memory occupied by mmap and memset | | 0G 440G 450G 460G 470G 480G 490G 500G | ====================+================================================+========= Execution time normalized to DRAM-only values | GEOMEAN --------------------+------------------------------------------------+--------- DRAM-only | 1.00 - - - - - - - | 1.00 CXL-only | 1.19 - - - - - - - | 1.19 default | - 1.00 1.05 1.08 1.12 1.14 1.18 1.18 | 1.11 DAMON tiered | - 1.03 1.03 1.03 1.03 1.03 1.07 *1.05 | 1.04 DAMON lazy | - 1.04 1.03 1.04 1.05 1.06 1.06 *1.06 | 1.05 ====================+================================================+========= CXL usage of redis-server in GB | AVERAGE --------------------+------------------------------------------------+--------- DRAM-only | 0.0 - - - - - - - | 0.0 CXL-only | 51.4 - - - - - - - | 51.4 default | - 0.6 10.6 20.5 30.5 40.5 47.6 50.4 | 28.7 DAMON tiered | - 0.6 0.5 0.4 0.7 0.8 7.1 5.6 | 2.2 DAMON lazy | - 0.5 3.0 4.5 5.4 6.4 9.4 9.1 | 5.5 ====================+================================================+========= Each test result is based on the execution environment as follows. DRAM-only: redis-server uses only local DRAM memory. CXL-only: redis-server uses only CXL memory. default: default memory policy(MPOL_DEFAULT). numa balancing disabled. DAMON tiered: DAMON enabled with DAMOS_MIGRATE_COLD for DRAM nodes and DAMOS_MIGRATE_HOT for CXL nodes. DAMON lazy: same as DAMON tiered, but turn on DAMON just before making memory access request via YCSB. The above result shows the "default" execution time goes up as the size of cold memory is increased from 440G to 500G because the more cold memory used, the more CXL memory is used for the target redis workload and this makes the execution time increase. However, "DAMON tiered" and other DAMON results show less slowdown because the DAMOS_MIGRATE_COLD action at DRAM node proactively demotes pre-allocated cold memory to CXL node and this free space at DRAM increases more chance to allocate hot or warm pages of redis-server to fast DRAM node. Moreover, DAMOS_MIGRATE_HOT action at CXL node also promotes hot pages of redis-server to DRAM node actively. As a result, it makes more memory of redis-server stay in DRAM node compared to "default" memory policy and this makes the performance improvement. Please note that the result numbers of "DAMON tiered" and "DAMON lazy" at 500G are marked with * stars, which means their test results are replaced with reproduced tests that didn't have OOM issue. That was needed because sometimes the test processes get OOM when DRAM has insufficient space. The DAMOS_MIGRATE_HOT doesn't kick reclaim but just gives up migration when there is not enough space at DRAM side. The problem happens when there is competition between normal allocation and migration and the migration is done before normal allocation, then the completely unrelated normal allocation can trigger reclaim, which incurs OOM. Because of this issue, I have also tested more cases with "demotion_enabled" flag enabled to make such reclaim doesn't trigger OOM, but just demote reclaimed pages. The following test results show more tests with "kswapd" marked. 2. YCSB zipfian distribution read only workload (with demotion_enabled true) memory pressure with cold memory on node0 with 512GB of local DRAM. ====================+================================================+========= | cold memory occupied by mmap and memset | | 0G 440G 450G 460G 470G 480G 490G 500G | ====================+================================================+========= Execution time normalized to DRAM-only values | GEOMEAN --------------------+------------------------------------------------+--------- DAMON tiered | - 1.03 1.03 1.03 1.03 1.03 1.07 1.05 | 1.04 DAMON lazy | - 1.04 1.03 1.04 1.05 1.06 1.06 1.06 | 1.05 DAMON tiered kswapd | - 1.03 1.03 1.03 1.03 1.02 1.02 1.03 | 1.03 DAMON lazy kswapd | - 1.04 1.04 1.04 1.03 1.05 1.04 1.05 | 1.04 ====================+================================================+========= CXL usage of redis-server in GB | AVERAGE --------------------+------------------------------------------------+--------- DAMON tiered | - 0.6 0.5 0.4 0.7 0.8 7.1 5.6 | 2.2 DAMON lazy | - 0.5 3.0 4.5 5.4 6.4 9.4 9.1 | 5.5 DAMON tiered kswapd | - 0.0 0.0 0.4 0.5 0.1 0.8 1.0 | 0.4 DAMON lazy kswapd | - 4.2 4.6 5.3 1.7 6.8 8.1 5.8 | 5.2 ====================+================================================+========= Each test result is based on the exeuction environment as follows. DAMON tiered: same as before DAMON lazy: same as before DAMON tiered kswapd: same as DAMON tiered, but turn on /sys/kernel/mm/numa/demotion_enabled to make kswapd or direct reclaim does demotion. DAMON lazy kswapd: same as DAMON lazy, but turn on /sys/kernel/mm/numa/demotion_enabled to make kswapd or direct reclaim does demotion. The "DAMON tiered kswapd" and "DAMON lazy kswapd" didn't trigger OOM at all unlike other tests because kswapd and direct reclaim from DRAM node can demote reclaimed pages to CXL node independently from DAMON actions and their results are slightly better than without having "demotion_enabled". In summary, the evaluation results show that DAMON memory management with DAMOS_MIGRATE_{HOT,COLD} actions reduces the performance slowdown compared to the "default" memory policy from 11% to 3~5% when the system runs with high memory pressure on its fast tier DRAM nodes. Having these DAMOS_MIGRATE_HOT and DAMOS_MIGRATE_COLD actions can make tiered memory systems run more efficiently under high memory pressures. This patch (of 7): The alloc_demote_folio can be used out of vmscan.c so it'd be better to remove static keyword from it. Link: https://lkml.kernel.org/r/20240614030010.751-1-honggyu.kim@sk.com Link: https://lkml.kernel.org/r/20240614030010.751-2-honggyu.kim@sk.comSigned-off-by:Honggyu Kim <honggyu.kim@sk.com> Reviewed-by:
SeongJae Park <sj@kernel.org> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org>
a00ce85a
