v2: style updates, weighted-interleave, rename partial-interleave to preferred-interleave, variety of bug fixes. --- This patch set is a proposal for set_mempolicy2 and get_mempolicy2 system calls. This is an extension to the existing mempolicy syscalls that allow for a more extensible mempolicy interface and new, complex memory policies. This RFC is broken into 4 patches for discussion: 1) A refactor of do_set_mempolicy that allows code reuse for the new syscalls when replacing the task mempolicy. 2) The implementation of get_mempolicy2 and set_mempolicy2 which includes a new uapi type: "struct mempolicy_args" and denotes the original mempolicies as "legacy". This allows the existing policies to be routed through the original interface. (note: only implemented on x86 at this time, though can be hacked into other architectures somewhat trivially) 3) The implementation of "preferred-interleave", a policy which applies a weight to the local node while interleaving. 4) The implementation of "weighted-interleave", a policy which applies weights to all enabled nodes while interleaving. x) Future Updates: ktest, numactl, and man page updates Besides the obvious proposal of extending the mempolicy subsystem for new policies, the core proposal is the addition of the new uapi type "struct mempolicy". In this proposal, the get and set interfaces use the same structure, and some fields may be ignored depending on the requested operation. This sample implementation of get_mempolicy allows for the retrieval of all information that would have previously required multiple calls to get_mempolicy, and implements an area for per-policy information. This allows for future extensibility, and would avoid the need for additional syscalls in the future. struct mempolicy_args { unsigned short mode; unsigned long *nodemask; unsigned long maxnode; unsigned short flags; struct { /* Memory allowed */ struct { unsigned long maxnode; unsigned long *nodemask; } allowed; /* Address information */ struct { unsigned long addr; unsigned long node; unsigned short mode; unsigned short flags; } addr; } get; union { /* Interleave */ struct { unsigned long next_node; /* get only */ } interleave; /* Preferred Interleave */ struct { unsigned long weight; /* get and set */ unsigned long next_node; /* get only */ } pil; /* Weighted Interleave */ struct { unsigned long next_node; /* get only */ unsigned char *weight; /* get and set */ } wil; }; }; In the third and fourth patch, we implement preferred and weighted interleave policies (respectively), which could not be implemented with the existing syscalls. We extend the internal mempolicy structure to include to include a new union area which can be used to host complex policy data. Example: union { /* Preferred Interleave: Allocate local count, then interleave */ struct { int weight; int count; } pil; /* Weighted Interleave */ struct { unsigned int il_weight; unsigned char cur_weight; unsigned char weights[MAX_NUMNODES]; } wil; }; Summary of Preferred Interleave: ================================ nodeset=0,1,2 interval=3 cpunode=0 The preferred node (cpunode) is the "preferred" node on which [weight] allocations are made before an interleave occurs. Over 10 consecutive allocations, the following nodes will be selected: [0,0,0,1,2,0,0,0,1,2] In this example, there is a 60%/20%/20% distribution of memory across the node set. This is a useful strategy if the goal is an even distribution of memory across all non-local nodes for the purpose of bandwidth AND task-node migrations are a possibility. In this case, the weight applies to whatever the local node happens to be at the time of the interleave, rather than a static node weight. Summary of Weighted Interleave: =============================== The weighted-interleave mempolicy implements weights per-node which are used to distribute memory while interleaving. For example: nodes: 0,1,2 weights: 5,3,2 Over 10 consecutive allocations, the following nodes will be selected: [0,0,0,0,0,1,1,1,2,2] If a node is enabled, the minimum weight is 1. If an enabled node ends up with a weight of 0 (cgroup updates can cause a runtime recalculation) a minimum of 1 is applied during interleave. This is a useful strategy if the goal is a non-even distribution of memory across a variety of nodes AND task-node migrations are NOT expected to occur (or the weights are approximately the same, relationally from all possible target nodes). This is because "Thread A" with weights set for best performance from the perspective of "Socket 0" may have a less-than-optimal interleave strategy if "Thread A" is migrated to "Socket 1". In this scenario, the bandwidth and latency attributes of each node will have changed, as will the local node. In the above example, a thread migrating from node 0 to node 1 will cause most of its memory to be allocated on remote nodes, which is less than optimal. Some notes for discussion ========================= 0) Why? In the coming age of CXL and a many-numa-node system with memory hosted on the PCIe bus, new memory policies are likely to be beneficial to experiment with and ultimately implement new allocation-time placement policies. Presently, much focus is placed on memory-usage monitoring and data migration, but these methods steal performance to accomplish what could be optimized for up-front. For example, if maximizing bandwith is preferable, then a statistical distribution of memory can be calculated fairly easily based on task location.. Getting a fair approximation of distribution at allocation can help reduce the migration load required after-the fact. This is the intent of the included preferred-interleave example, which allows for an approximate distribution of memory, where the local node is still the preferred location for the majority of memory. 1) Maybe this should be a set of sysfs interfaces? This would involve adding a /proc/pid/mempolicy interface that allows for external processes to interrogate and change the mempolicy of running processes. This would be a fundamental change to the mempolicy subsystem. I attempted this, but eventually came to the conclusion that it would require a much more radical re-write of mempolicy.c code due concurrency issues. Notably, mempolicy.c is very "current"--centric, and is not well designed for runtime changes to nodemask (and, subsequently, the new weights added to struct mempolicy). I avoided that for this RFC as it seemed far more radical than proposing a set/get_mempolicy2 interface. Though technically it could be done. 2) Why not do this in cgroups or memtier? Both have the issue of functionally being a "global" setting, in the sense that cgroups/memtier implemented weights would produce poor results on processes whose threads span multiple sockets (or on a thread migration). Consider the following scenario: Node 0 - Socket 0 DRAM Node 1 - Socket 1 DRAM Node 2 - Socket 0 local CXL Node 3 - Socket 1 local CXL Weights: [0:4, 1:2, 2:2, 3:1] The "Tiers" in this case are essentially [0, 1-2, 3] We have 2 tasks in our cgroup: Thread A - socket 0 Thread B - socket 1 In this scenario, Thread B will have a very poor distribution of memory, with most of its memory landing on a remote-socket. Instead, it's preferable for workloads to stick to a single socket where possible, and future work will need to be done to determine how to handle workloads which span sockets. Due to the above mentioned issues with concurrency, this may be quite some time. In the meantime, there is user for weights to be carried per-task. For migrations: Weights could be recalculated based on the new location of the task. This recalculation of weights is not included in this patch set, but could be done as an extension to weighted interleave, where a thread that detects it has been migrated works with memtier.c to adjust its weights internally. So basically even if you implement these things in cgroups/memtier, you still require per-task information (local node) to adjust the weights. My proposal: Just do it in mempolicy and use things like cgroups/memtier to enrich that implementation, rather than the other way around. 3) Do we need this level extensibility? Presently the ability to dictate allocation-time placement is limited to a few primitive mechanisms: 1) existing mempolicy, and those that can be implemented using the existing interface. 2) numa-aware applications, requiring code changes. 3) LDPRELOAD methods, which have compability issues. For the sake of compatibility, being able to extend numactl to include newer, more complex policies would be beneficial. Gregory Price (4): mm/mempolicy: refactor do_set_mempolicy for code re-use mm/mempolicy: Implement set_mempolicy2 and get_mempolicy2 syscalls mm/mempolicy: implement a preferred-interleave mempolicy mm/mempolicy: implement a weighted-interleave mempolicy arch/x86/entry/syscalls/syscall_32.tbl | 2 + arch/x86/entry/syscalls/syscall_64.tbl | 2 + include/linux/mempolicy.h | 14 + include/linux/syscalls.h | 4 + include/uapi/asm-generic/unistd.h | 10 +- include/uapi/linux/mempolicy.h | 41 ++ mm/mempolicy.c | 688 ++++++++++++++++++++++++- 7 files changed, 741 insertions(+), 20 deletions(-) -- 2.39.1