- 09 Oct, 2013 40 commits
-
-
Peter Zijlstra authored
This patch classifies scheduler domains and runqueues into types depending the number of tasks that are about their NUMA placement and the number that are currently running on their preferred node. The types are regular: There are tasks running that do not care about their NUMA placement. remote: There are tasks running that care about their placement but are currently running on a node remote to their ideal placement all: No distinction To implement this the patch tracks the number of tasks that are optimally NUMA placed (rq->nr_preferred_running) and the number of tasks running that care about their placement (nr_numa_running). The load balancer uses this information to avoid migrating idea placed NUMA tasks as long as better options for load balancing exists. For example, it will not consider balancing between a group whose tasks are all perfectly placed and a group with remote tasks. Signed-off-by:
Peter Zijlstra <peterz@infradead.org> Signed-off-by:
Mel Gorman <mgorman@suse.de> Reviewed-by:
Rik van Riel <riel@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Link: http://lkml.kernel.org/r/1381141781-10992-56-git-send-email-mgorman@suse.deSigned-off-by:
Ingo Molnar <mingo@kernel.org>
-
Rik van Riel authored
This patch separately considers task and group affinities when searching for swap candidates during NUMA placement. If tasks are part of the same group, or no group at all, the task weights are considered. Some hysteresis is added to prevent tasks within one group from getting bounced between NUMA nodes due to tiny differences. If tasks are part of different groups, the code compares group weights, in order to favor grouping task groups together. The patch also changes the group weight multiplier to be the same as the task weight multiplier, since the two are no longer added up like before. Signed-off-by:
-
Rik van Riel authored
This patch separately considers task and group affinities when searching for swap candidates during task NUMA placement. If tasks are not part of a group or the same group then the task weights are considered. Otherwise the group weights are compared. Signed-off-by:
-
Ingo Molnar authored
Signed-off-by:
-
Mel Gorman authored
Having multiple tasks in a group go through task_numa_placement simultaneously can lead to a task picking a wrong node to run on, because the group stats may be in the middle of an update. This patch avoids parallel updates by holding the numa_group lock during placement decisions. Signed-off-by:
-
Rik van Riel authored
It is possible for a task in a numa group to call exec, and have the new (unrelated) executable inherit the numa group association from its former self. This has the potential to break numa grouping, and is trivial to fix. Signed-off-by:
-
Mel Gorman authored
This patch uses the fraction of faults on a particular node for both task and group, to figure out the best node to place a task. If the task and group statistics disagree on what the preferred node should be then a full rescan will select the node with the best combined weight. Signed-off-by:
-
Rik van Riel authored
A newly spawned thread inside a process should stay on the same NUMA node as its parent. This prevents processes from being "torn" across multiple NUMA nodes every time they spawn a new thread. Signed-off-by:
-
Mel Gorman authored
With the THP migration races closed it is still possible to occasionally see corruption. The problem is related to handling PMD pages in batch. When a page fault is handled it can be assumed that the page being faulted will also be flushed from the TLB. The same flushing does not happen when handling PMD pages in batch. Fixing is straight forward but there are a number of reasons not to 1. Multiple TLB flushes may have to be sent depending on what pages get migrated 2. The handling of PMDs in batch means that faults get accounted to the task that is handling the fault. While care is taken to only mark PMDs where the last CPU and PID match it can still have problems due to PID truncation when matching PIDs. 3. Batching on the PMD level may reduce faults but setting pmd_numa requires taking a heavy lock that can contend with THP migration and handling the fault requires the release/acquisition of the PTL for every page migrated. It's still pretty heavy. PMD batch handling is not something that people ever have been happy with. This patch removes it and later patches will deal with the additional fault overhead using more installigent migrate rate adaption. Signed-off-by:
-
Peter Zijlstra authored
And here's a little something to make sure not the whole world ends up in a single group. As while we don't migrate shared executable pages, we do scan/fault on them. And since everybody links to libc, everybody ends up in the same group. Suggested-by:
-
Rik van Riel authored
After page migration, the new page has the nidpid unset. This makes every fault on a recently migrated page look like a first numa fault, leading to another page migration. Copying over the nidpid at page migration time should prevent erroneous migrations of recently migrated pages. Signed-off-by:
-
Mel Gorman authored
It is desirable to model from userspace how the scheduler groups tasks over time. This patch adds an ID to the numa_group and reports it via /proc/PID/status. Signed-off-by:
-
Peter Zijlstra authored
While parallel applications tend to align their data on the cache boundary, they tend not to align on the page or THP boundary. Consequently tasks that partition their data can still "false-share" pages presenting a problem for optimal NUMA placement. This patch uses NUMA hinting faults to chain tasks together into numa_groups. As well as storing the NID a task was running on when accessing a page a truncated representation of the faulting PID is stored. If subsequent faults are from different PIDs it is reasonable to assume that those two tasks share a page and are candidates for being grouped together. Note that this patch makes no scheduling decisions based on the grouping information. Signed-off-by:
-
Peter Zijlstra authored
Change the per page last fault tracking to use cpu,pid instead of nid,pid. This will allow us to try and lookup the alternate task more easily. Note that even though it is the cpu that is store in the page flags that the mpol_misplaced decision is still based on the node. Signed-off-by:
-
Rik van Riel authored
The load balancer will spread workloads across multiple NUMA nodes, in order to balance the load on the system. This means that sometimes a task's preferred node has available capacity, but moving the task there will not succeed, because that would create too large an imbalance. In that case, other NUMA nodes need to be considered. Signed-off-by:
-
Mel Gorman authored
A tasks preferred node is selected based on the number of faults recorded for a node but the actual task_numa_migate() conducts a global search regardless of the preferred nid. This patch checks if the preferred nid has capacity and if so, searches for a CPU within that node. This avoids a global search when the preferred node is not overloaded. Signed-off-by:
-
Mel Gorman authored
This patch implements a system-wide search for swap/migration candidates based on total NUMA hinting faults. It has a balance limit, however it doesn't properly consider total node balance. In the old scheme a task selected a preferred node based on the highest number of private faults recorded on the node. In this scheme, the preferred node is based on the total number of faults. If the preferred node for a task changes then task_numa_migrate will search the whole system looking for tasks to swap with that would improve both the overall compute balance and minimise the expected number of remote NUMA hinting faults. Not there is no guarantee that the node the source task is placed on by task_numa_migrate() has any relationship to the newly selected task->numa_preferred_nid due to compute overloading. Signed-off-by:
-
Peter Zijlstra authored
Use the new stop_two_cpus() to implement migrate_swap(), a function that flips two tasks between their respective cpus. I'm fairly sure there's a less crude way than employing the stop_two_cpus() method, but everything I tried either got horribly fragile and/or complex. So keep it simple for now. The notable detail is how we 'migrate' tasks that aren't runnable anymore. We'll make it appear like we migrated them before they went to sleep. The sole difference is the previous cpu in the wakeup path, so we override this. Signed-off-by:
-
Peter Zijlstra authored
Introduce stop_two_cpus() in order to allow controlled swapping of two tasks. It repurposes the stop_machine() state machine but only stops the two cpus which we can do with on-stack structures and avoid machine wide synchronization issues. The ordering of CPUs is important to avoid deadlocks. If unordered then two cpus calling stop_two_cpus on each other simultaneously would attempt to queue in the opposite order on each CPU causing an AB-BA style deadlock. By always having the lowest number CPU doing the queueing of works, we can guarantee that works are always queued in the same order, and deadlocks are avoided. Signed-off-by:
-
Mel Gorman authored
Base page PMD faulting is meant to batch handle NUMA hinting faults from PTEs. However, even is no PTE faults would ever be handled within a range the kernel still traps PMD hinting faults. This patch avoids the overhead. Signed-off-by:
-
Mel Gorman authored
NUMA hinting faults will not migrate a shared executable page mapped by multiple processes on the grounds that the data is probably in the CPU cache already and the page may just bounce between tasks running on multipl nodes. Even if the migration is avoided, there is still the overhead of trapping the fault, updating the statistics, making scheduler placement decisions based on the information etc. If we are never going to migrate the page, it is overhead for no gain and worse a process may be placed on a sub-optimal node for shared executable pages. This patch avoids trapping faults for shared libraries entirely. Signed-off-by:
-
Rik van Riel authored
When a task is already running on its preferred node, increment numa_migrate_seq to indicate that the task is settled if migration is temporarily disabled, and memory should migrate towards it. Signed-off-by:
-
Mel Gorman authored
When a preferred node is selected for a tasks there is an attempt to migrate the task to a CPU there. This may fail in which case the task will only migrate if the active load balancer takes action. This may never happen if the conditions are not right. This patch will check at NUMA hinting fault time if another attempt should be made to migrate the task. It will only make an attempt once every five seconds. Signed-off-by:
-
Mel Gorman authored
This patch replaces find_idlest_cpu_node with task_numa_find_cpu. find_idlest_cpu_node has two critical limitations. It does not take the scheduling class into account when calculating the load and it is unsuitable for using when comparing loads between NUMA nodes. task_numa_find_cpu uses similar load calculations to wake_affine() when selecting the least loaded CPU within a scheduling domain common to the source and destimation nodes. It avoids causing CPU load imbalances in the machine by refusing to migrate if the relative load on the target CPU is higher than the source CPU. Signed-off-by:
-
Mel Gorman authored
There is a 90% regression observed with a large Oracle performance test on a 4 node system. Profiles indicated that the overhead was due to contention on sp_lock when looking up shared memory policies. These policies do not have the appropriate flags to allow them to be automatically balanced so trapping faults on them is pointless. This patch skips VMAs that do not have MPOL_F_MOF set. [riel@redhat.com: Initial patch] Signed-off-by:
Joe Mario <jmario@redhat.com> Reviewed-by:
-
Rik van Riel authored
The load balancer can move tasks between nodes and does not take NUMA locality into account. With automatic NUMA balancing this may result in the tasks working set being migrated to the new node. However, as the fault buffer will still store faults from the old node the schduler may decide to reset the preferred node and migrate the task back resulting in more migrations. The ideal would be that the scheduler did not migrate tasks with a heavy memory footprint but this may result nodes being overloaded. We could also discard the fault information on task migration but this would still cause all the tasks working set to be migrated. This patch simply avoids migrating the memory for a short time after a task is migrated. Signed-off-by:
-
Mel Gorman authored
Ideally it would be possible to distinguish between NUMA hinting faults that are private to a task and those that are shared. If treated identically there is a risk that shared pages bounce between nodes depending on the order they are referenced by tasks. Ultimately what is desirable is that task private pages remain local to the task while shared pages are interleaved between sharing tasks running on different nodes to give good average performance. This is further complicated by THP as even applications that partition their data may not be partitioning on a huge page boundary. To start with, this patch assumes that multi-threaded or multi-process applications partition their data and that in general the private accesses are more important for cpu->memory locality in the general case. Also, no new infrastructure is required to treat private pages properly but interleaving for shared pages requires additional infrastructure. To detect private accesses the pid of the last accessing task is required but the storage requirements are a high. This patch borrows heavily from Ingo Molnar's patch "numa, mm, sched: Implement last-CPU+PID hash tracking" to encode some bits from the last accessing task in the page flags as well as the node information. Collisions will occur but it is better than just depending on the node information. Node information is then used to determine if a page needs to migrate. The PID information is used to detect private/shared accesses. The preferred NUMA node is selected based on where the maximum number of approximately private faults were measured. Shared faults are not taken into consideration for a few reasons. First, if there are many tasks sharing the page then they'll all move towards the same node. The node will be compute overloaded and then scheduled away later only to bounce back again. Alternatively the shared tasks would just bounce around nodes because the fault information is effectively noise. Either way accounting for shared faults the same as private faults can result in lower performance overall. The second reason is based on a hypothetical workload that has a small number of very important, heavily accessed private pages but a large shared array. The shared array would dominate the number of faults and be selected as a preferred node even though it's the wrong decision. The third reason is that multiple threads in a process will race each other to fault the shared page making the fault information unreliable. Signed-off-by:
-
Mel Gorman authored
task_numa_work skips small VMAs. At the time the logic was to reduce the scanning overhead which was considerable. It is a dubious hack at best. It would make much more sense to cache where faults have been observed and only rescan those regions during subsequent PTE scans. Remove this hack as motivation to do it properly in the future. Signed-off-by:
-
Mel Gorman authored
Currently automatic NUMA balancing is unable to distinguish between false shared versus private pages except by ignoring pages with an elevated page_mapcount entirely. This avoids shared pages bouncing between the nodes whose task is using them but that is ignored quite a lot of data. This patch kicks away the training wheels in preparation for adding support for identifying shared/private pages is now in place. The ordering is so that the impact of the shared/private detection can be easily measured. Note that the patch does not migrate shared, file-backed within vmas marked VM_EXEC as these are generally shared library pages. Migrating such pages is not beneficial as there is an expectation they are read-shared between caches and iTLB and iCache pressure is generally low. Signed-off-by:
-
Mel Gorman authored
task_numa_placement checks current->mm but after buffers for faults have already been uselessly allocated. Move the check earlier. [peterz@infradead.org: Identified the problem] Signed-off-by:
-
Mel Gorman authored
Ideally it would be possible to distinguish between NUMA hinting faults that are private to a task and those that are shared. This patch prepares infrastructure for separately accounting shared and private faults by allocating the necessary buffers and passing in relevant information. For now, all faults are treated as private and detection will be introduced later. Signed-off-by:
-
Mel Gorman authored
A preferred node is selected based on the node the most NUMA hinting faults was incurred on. There is no guarantee that the task is running on that node at the time so this patch rescheules the task to run on the most idle CPU of the selected node when selected. This avoids waiting for the balancer to make a decision. Signed-off-by:
-
Mel Gorman authored
Just as "sched: Favour moving tasks towards the preferred node" favours moving tasks towards nodes with a higher number of recorded NUMA hinting faults, this patch resists moving tasks towards nodes with lower faults. Signed-off-by:
-
Mel Gorman authored
This patch favours moving tasks towards NUMA node that recorded a higher number of NUMA faults during active load balancing. Ideally this is self-reinforcing as the longer the task runs on that node, the more faults it should incur causing task_numa_placement to keep the task running on that node. In reality a big weakness is that the nodes CPUs can be overloaded and it would be more efficient to queue tasks on an idle node and migrate to the new node. This would require additional smarts in the balancer so for now the balancer will simply prefer to place the task on the preferred node for a PTE scans which is controlled by the numa_balancing_settle_count sysctl. Once the settle_count number of scans has complete the schedule is free to place the task on an alternative node if the load is imbalanced. [srikar@linux.vnet.ibm.com: Fixed statistics] Signed-off-by:
-
Mel Gorman authored
NUMA hinting fault counts and placement decisions are both recorded in the same array which distorts the samples in an unpredictable fashion. The values linearly accumulate during the scan and then decay creating a sawtooth-like pattern in the per-node counts. It also means that placement decisions are time sensitive. At best it means that it is very difficult to state that the buffer holds a decaying average of past faulting behaviour. At worst, it can confuse the load balancer if it sees one node with an artifically high count due to very recent faulting activity and may create a bouncing effect. This patch adds a second array. numa_faults stores the historical data which is used for placement decisions. numa_faults_buffer holds the fault activity during the current scan window. When the scan completes, numa_faults decays and the values from numa_faults_buffer are copied across. Signed-off-by:
-
Mel Gorman authored
This patch selects a preferred node for a task to run on based on the NUMA hinting faults. This information is later used to migrate tasks towards the node during balancing. Signed-off-by:
-
Mel Gorman authored
This patch tracks what nodes numa hinting faults were incurred on. This information is later used to schedule a task on the node storing the pages most frequently faulted by the task. Signed-off-by:
-
Mel Gorman authored
NUMA PTE scanning slows if a NUMA hinting fault was trapped and no page was migrated. For long-lived but idle processes there may be no faults but the scan rate will be high and just waste CPU. This patch will slow the scan rate for processes that are not trapping faults. Signed-off-by:
-
Mel Gorman authored
The NUMA PTE scan rate is controlled with a combination of the numa_balancing_scan_period_min, numa_balancing_scan_period_max and numa_balancing_scan_size. This scan rate is independent of the size of the task and as an aside it is further complicated by the fact that numa_balancing_scan_size controls how many pages are marked pte_numa and not how much virtual memory is scanned. In combination, it is almost impossible to meaningfully tune the min and max scan periods and reasoning about performance is complex when the time to complete a full scan is is partially a function of the tasks memory size. This patch alters the semantic of the min and max tunables to be about tuning the length time it takes to complete a scan of a tasks occupied virtual address space. Conceptually this is a lot easier to understand. There is a "sanity" check to ensure the scan rate is never extremely fast based on the amount of virtual memory that should be scanned in a second. The default of 2.5G seems arbitrary but it is to have the maximum scan rate after the patch roughly match the maximum scan rate before the patch was applied. On a similar note, numa_scan_period is in milliseconds and not jiffies. Properly placed pages slow the scanning rate but adding 10 jiffies to numa_scan_period means that the rate scanning slows depends on HZ which is confusing. Get rid of the jiffies_to_msec conversion and treat it as ms. Signed-off-by:
-
Mel Gorman authored
Scan delay logic and resets are currently initialised to start scanning immediately instead of delaying properly. Initialise them properly at fork time and catch when a new mm has been allocated. Signed-off-by:
-
