| Message ID | 20210426065946.40491-3-wuyun.abel@bytedance.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | cgroup2: introduce cpuset.mems.migration | expand |
Hello, On Mon, Apr 26, 2021 at 02:59:45PM +0800, Abel Wu wrote: > When a NUMA node is assigned to numa-service, the workload > on that node needs to be moved away fast and complete. The > main aspects we cared about on the eviction are as follows: > > a) it should complete soon enough so that numa-services > won’t wait too long to hurt user experience > b) the workloads to be evicted could have massive usage on > memory, and migrating such amount of memory may lead to > a sudden severe performance drop lasting tens of seconds > that some certain workloads may not afford > c) the impact of the eviction should be limited within the > source and destination nodes > d) cgroup interface is preferred > > So we come to a thought that: > > 1) fire up numa-services without waiting for memory migration > 2) memory migration can be done asynchronously by using spare > memory bandwidth > > AutoNUMA seems to be a solution, but its scope is global which > violates c&d. And cpuset.memory_migrate performs in a synchronous I don't think d) in itself is a valid requirement. How does it violate c)? > fashion which breaks a&b. So a mixture of them, the new cgroup2 > interface cpuset.mems.migration, is introduced. > > The new cpuset.mems.migration supports three modes: > > - "none" mode, meaning migration disabled > - "sync" mode, which is exactly the same as the cgroup v1 > interface cpuset.memory_migrate > - "lazy" mode, when walking through all the pages, unlike > cpuset.memory_migrate, it only sets pages to protnone, > and numa faults triggered by later touch will handle the > movement. cpuset is already involved in NUMA allocation but it always felt like something bolted on - it's weird to have cpu to NUMA node settings at global level and then to have possibly conflicting direct NUMA configuration via cpuset. My preference would be putting as much configuration as possible on the mm / autonuma side and let cpuset's node confinements further restrict their operations rather than cpuset having its own set of policy configurations. Johannes, what are your thoughts? Thanks.
Hello Tejun, thanks for your review, On 4/27/21 10:43 PM, Tejun Heo wrote: > Hello, > > On Mon, Apr 26, 2021 at 02:59:45PM +0800, Abel Wu wrote: >> When a NUMA node is assigned to numa-service, the workload >> on that node needs to be moved away fast and complete. The >> main aspects we cared about on the eviction are as follows: >> >> a) it should complete soon enough so that numa-services >> won’t wait too long to hurt user experience >> b) the workloads to be evicted could have massive usage on >> memory, and migrating such amount of memory may lead to >> a sudden severe performance drop lasting tens of seconds >> that some certain workloads may not afford >> c) the impact of the eviction should be limited within the >> source and destination nodes >> d) cgroup interface is preferred >> >> So we come to a thought that: >> >> 1) fire up numa-services without waiting for memory migration >> 2) memory migration can be done asynchronously by using spare >> memory bandwidth >> >> AutoNUMA seems to be a solution, but its scope is global which >> violates c&d. And cpuset.memory_migrate performs in a synchronous > > I don't think d) in itself is a valid requirement. How does it violate c)? Yes, d) is more like a preference, since we operate in cgroup level. Process/thread level interfaces are also acceptable. AutoNUMA violates c) in its global effect that not only the source and destination nodes, the processes running on other nodes would also suffer from unwanted overhead due to numa faults. And besides the global effect, one-shot mode migration is expected in this scenario, like cpuset.memory_migrate, rather than autonuma's periodic behavior. > >> fashion which breaks a&b. So a mixture of them, the new cgroup2 >> interface cpuset.mems.migration, is introduced. >> >> The new cpuset.mems.migration supports three modes: >> >> - "none" mode, meaning migration disabled >> - "sync" mode, which is exactly the same as the cgroup v1 >> interface cpuset.memory_migrate >> - "lazy" mode, when walking through all the pages, unlike >> cpuset.memory_migrate, it only sets pages to protnone, >> and numa faults triggered by later touch will handle the >> movement. > > cpuset is already involved in NUMA allocation but it always felt like > something bolted on - it's weird to have cpu to NUMA node settings at global > level and then to have possibly conflicting direct NUMA configuration via > cpuset. My preference would be putting as much configuration as possible on > the mm / autonuma side and let cpuset's node confinements further restrict > their operations rather than cpuset having its own set of policy > configurations. Such conflicting configuration exists in our system in order to reduce TCO, and yet we haven't found a proper way to get rid of it. Say a numa-service occupies the whole memory of a node but still leave some cpus free. The free cpus may be assigned to some services, the ones that are not memory latency sensitive and are forbidden to use local memory. Thoughts? Thanks, Abel > > Johannes, what are your thoughts? > > Thanks. >
ping :) On 4/27/21 10:43 PM, Tejun Heo wrote: > Hello, > > On Mon, Apr 26, 2021 at 02:59:45PM +0800, Abel Wu wrote: >> When a NUMA node is assigned to numa-service, the workload >> on that node needs to be moved away fast and complete. The >> main aspects we cared about on the eviction are as follows: >> >> a) it should complete soon enough so that numa-services >> won’t wait too long to hurt user experience >> b) the workloads to be evicted could have massive usage on >> memory, and migrating such amount of memory may lead to >> a sudden severe performance drop lasting tens of seconds >> that some certain workloads may not afford >> c) the impact of the eviction should be limited within the >> source and destination nodes >> d) cgroup interface is preferred >> >> So we come to a thought that: >> >> 1) fire up numa-services without waiting for memory migration >> 2) memory migration can be done asynchronously by using spare >> memory bandwidth >> >> AutoNUMA seems to be a solution, but its scope is global which >> violates c&d. And cpuset.memory_migrate performs in a synchronous > > I don't think d) in itself is a valid requirement. How does it violate c)? > >> fashion which breaks a&b. So a mixture of them, the new cgroup2 >> interface cpuset.mems.migration, is introduced. >> >> The new cpuset.mems.migration supports three modes: >> >> - "none" mode, meaning migration disabled >> - "sync" mode, which is exactly the same as the cgroup v1 >> interface cpuset.memory_migrate >> - "lazy" mode, when walking through all the pages, unlike >> cpuset.memory_migrate, it only sets pages to protnone, >> and numa faults triggered by later touch will handle the >> movement. > > cpuset is already involved in NUMA allocation but it always felt like > something bolted on - it's weird to have cpu to NUMA node settings at global > level and then to have possibly conflicting direct NUMA configuration via > cpuset. My preference would be putting as much configuration as possible on > the mm / autonuma side and let cpuset's node confinements further restrict > their operations rather than cpuset having its own set of policy > configurations. > > Johannes, what are your thoughts? > > Thanks. >
On Tue, Apr 27, 2021 at 10:43:31AM -0400, Tejun Heo wrote: > Hello, > > On Mon, Apr 26, 2021 at 02:59:45PM +0800, Abel Wu wrote: > > When a NUMA node is assigned to numa-service, the workload > > on that node needs to be moved away fast and complete. The > > main aspects we cared about on the eviction are as follows: > > > > a) it should complete soon enough so that numa-services > > won’t wait too long to hurt user experience > > b) the workloads to be evicted could have massive usage on > > memory, and migrating such amount of memory may lead to > > a sudden severe performance drop lasting tens of seconds > > that some certain workloads may not afford > > c) the impact of the eviction should be limited within the > > source and destination nodes > > d) cgroup interface is preferred > > > > So we come to a thought that: > > > > 1) fire up numa-services without waiting for memory migration > > 2) memory migration can be done asynchronously by using spare > > memory bandwidth > > > > AutoNUMA seems to be a solution, but its scope is global which > > violates c&d. And cpuset.memory_migrate performs in a synchronous > > I don't think d) in itself is a valid requirement. How does it violate c)? > > > fashion which breaks a&b. So a mixture of them, the new cgroup2 > > interface cpuset.mems.migration, is introduced. > > > > The new cpuset.mems.migration supports three modes: > > > > - "none" mode, meaning migration disabled > > - "sync" mode, which is exactly the same as the cgroup v1 > > interface cpuset.memory_migrate > > - "lazy" mode, when walking through all the pages, unlike > > cpuset.memory_migrate, it only sets pages to protnone, > > and numa faults triggered by later touch will handle the > > movement. > > cpuset is already involved in NUMA allocation but it always felt like > something bolted on - it's weird to have cpu to NUMA node settings at global > level and then to have possibly conflicting direct NUMA configuration via > cpuset. My preference would be putting as much configuration as possible on > the mm / autonuma side and let cpuset's node confinements further restrict > their operations rather than cpuset having its own set of policy > configurations. > > Johannes, what are your thoughts? This is basically a cgroup interface for the existing MPOL_MF_LAZY / MPOL_F_MOF flags, which are per task (set_mempolicy()) and per-vma (mbind()) scope respectively. They're not per-node, so cannot be cgroupified through cpuset's node restrictions alone, and I understand why a cgroup interface could be convenient. On the other hand, this is not really about configuring a shared resource. Rather it's using cgroups to set an arbitrary task parameter on a bunch of tasks simultaneously. It's the SIMD type usecase of cgroup1 that we tried to get away from in cgroup2, simply because it's so unbounded in scope. There are *a lot* of possible task parameters, and we could add a lot of kernel interfaces that boil down to css_task_iter and setting or clearing a task flag. So I'm also thinking this cgroup interface isn't desirable. If you want to control numa policies of tasks from the outside, it's probably best to extend the numa syscall interface to work on pids. And then use cgroup.procs to cgroupify the operation from userspace. Or extend the NUMA interface to make the system-wide default behavior configurable, so that you can set MPOL_F_MOF in there (without having to enable autonuma). But yeah, cgroups doesn't seem like the right place to do this. Thanks
diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c index a945504c0ae7..ee84f168eea8 100644 --- a/kernel/cgroup/cpuset.c +++ b/kernel/cgroup/cpuset.c @@ -212,6 +212,7 @@ typedef enum { CS_MEM_EXCLUSIVE, CS_MEM_HARDWALL, CS_MEMORY_MIGRATE, + CS_MEMORY_MIGRATE_LAZY, CS_SCHED_LOAD_BALANCE, CS_SPREAD_PAGE, CS_SPREAD_SLAB, @@ -248,6 +249,11 @@ static inline int is_memory_migrate(const struct cpuset *cs) return test_bit(CS_MEMORY_MIGRATE, &cs->flags); } +static inline int is_memory_migrate_lazy(const struct cpuset *cs) +{ + return test_bit(CS_MEMORY_MIGRATE_LAZY, &cs->flags); +} + static inline int is_spread_page(const struct cpuset *cs) { return test_bit(CS_SPREAD_PAGE, &cs->flags); @@ -1594,6 +1600,7 @@ struct cpuset_migrate_mm_work { struct mm_struct *mm; nodemask_t from; nodemask_t to; + int flags; }; static void cpuset_migrate_mm_workfn(struct work_struct *work) @@ -1602,21 +1609,29 @@ static void cpuset_migrate_mm_workfn(struct work_struct *work) container_of(work, struct cpuset_migrate_mm_work, work); /* on a wq worker, no need to worry about %current's mems_allowed */ - do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL); + do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, mwork->flags); mmput(mwork->mm); kfree(mwork); } -static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from, - const nodemask_t *to) +static void cpuset_migrate_mm(struct cpuset *cs, struct mm_struct *mm, + const nodemask_t *from, const nodemask_t *to) { - struct cpuset_migrate_mm_work *mwork; + struct cpuset_migrate_mm_work *mwork = NULL; + int flags = 0; - mwork = kzalloc(sizeof(*mwork), GFP_KERNEL); + if (is_memory_migrate_lazy(cs)) + flags = MPOL_MF_LAZY; + else if (is_memory_migrate(cs)) + flags = MPOL_MF_MOVE_ALL; + + if (flags) + mwork = kzalloc(sizeof(*mwork), GFP_KERNEL); if (mwork) { mwork->mm = mm; mwork->from = *from; mwork->to = *to; + mwork->flags = flags; INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn); queue_work(cpuset_migrate_mm_wq, &mwork->work); } else { @@ -1690,7 +1705,6 @@ static void update_tasks_nodemask(struct cpuset *cs) css_task_iter_start(&cs->css, 0, &it); while ((task = css_task_iter_next(&it))) { struct mm_struct *mm; - bool migrate; cpuset_change_task_nodemask(task, &newmems); @@ -1698,13 +1712,8 @@ static void update_tasks_nodemask(struct cpuset *cs) if (!mm) continue; - migrate = is_memory_migrate(cs); - mpol_rebind_mm(mm, &cs->mems_allowed); - if (migrate) - cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems); - else - mmput(mm); + cpuset_migrate_mm(cs, mm, &cs->old_mems_allowed, &newmems); } css_task_iter_end(&it); @@ -1911,6 +1920,11 @@ static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, else clear_bit(bit, &trialcs->flags); + if (bit == CS_MEMORY_MIGRATE) + clear_bit(CS_MEMORY_MIGRATE_LAZY, &trialcs->flags); + if (bit == CS_MEMORY_MIGRATE_LAZY) + clear_bit(CS_MEMORY_MIGRATE, &trialcs->flags); + err = validate_change(cs, trialcs); if (err < 0) goto out; @@ -2237,11 +2251,8 @@ static void cpuset_attach(struct cgroup_taskset *tset) * @old_mems_allowed is the right nodesets that we * migrate mm from. */ - if (is_memory_migrate(cs)) - cpuset_migrate_mm(mm, &oldcs->old_mems_allowed, - &cpuset_attach_nodemask_to); - else - mmput(mm); + cpuset_migrate_mm(cs, mm, &oldcs->old_mems_allowed, + &cpuset_attach_nodemask_to); } } @@ -2258,6 +2269,7 @@ static void cpuset_attach(struct cgroup_taskset *tset) typedef enum { FILE_MEMORY_MIGRATE, + FILE_MEMORY_MIGRATE_LAZY, FILE_CPULIST, FILE_MEMLIST, FILE_EFFECTIVE_CPULIST, @@ -2275,11 +2287,8 @@ typedef enum { FILE_SPREAD_SLAB, } cpuset_filetype_t; -static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft, - u64 val) +static int __cpuset_write_u64(struct cpuset *cs, cpuset_filetype_t type, u64 val) { - struct cpuset *cs = css_cs(css); - cpuset_filetype_t type = cft->private; int retval = 0; get_online_cpus(); @@ -2305,6 +2314,9 @@ static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft, case FILE_MEMORY_MIGRATE: retval = update_flag(CS_MEMORY_MIGRATE, cs, val); break; + case FILE_MEMORY_MIGRATE_LAZY: + retval = update_flag(CS_MEMORY_MIGRATE_LAZY, cs, val); + break; case FILE_MEMORY_PRESSURE_ENABLED: cpuset_memory_pressure_enabled = !!val; break; @@ -2324,6 +2336,12 @@ static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft, return retval; } +static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft, + u64 val) +{ + return __cpuset_write_u64(css_cs(css), cft->private, val); +} + static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft, s64 val) { @@ -2473,6 +2491,8 @@ static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft) return is_sched_load_balance(cs); case FILE_MEMORY_MIGRATE: return is_memory_migrate(cs); + case FILE_MEMORY_MIGRATE_LAZY: + return is_memory_migrate_lazy(cs); case FILE_MEMORY_PRESSURE_ENABLED: return cpuset_memory_pressure_enabled; case FILE_MEMORY_PRESSURE: @@ -2555,6 +2575,40 @@ static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf, return retval ?: nbytes; } +static int cpuset_mm_migration_show(struct seq_file *seq, void *v) +{ + struct cpuset *cs = css_cs(seq_css(seq)); + + if (is_memory_migrate_lazy(cs)) + seq_puts(seq, "lazy\n"); + else if (is_memory_migrate(cs)) + seq_puts(seq, "sync\n"); + else + seq_puts(seq, "none\n"); + return 0; +} + +static ssize_t cpuset_mm_migration_write(struct kernfs_open_file *of, + char *buf, size_t nbytes, loff_t off) +{ + struct cpuset *cs = css_cs(of_css(of)); + cpuset_filetype_t type = FILE_MEMORY_MIGRATE; + int turning_on = 1; + int retval; + + buf = strstrip(buf); + + if (!strcmp(buf, "none")) + turning_on = 0; + else if (!strcmp(buf, "lazy")) + type = FILE_MEMORY_MIGRATE_LAZY; + else if (strcmp(buf, "sync")) + return -EINVAL; + + retval = __cpuset_write_u64(cs, type, turning_on); + return retval ?: nbytes; +} + /* * for the common functions, 'private' gives the type of file */ @@ -2711,6 +2765,14 @@ static struct cftype dfl_files[] = { .flags = CFTYPE_DEBUG, }, + { + .name = "mems.migration", + .seq_show = cpuset_mm_migration_show, + .write = cpuset_mm_migration_write, + .private = FILE_MEMORY_MIGRATE, + .flags = CFTYPE_NOT_ON_ROOT, + }, + { } /* terminate */ }; diff --git a/mm/mempolicy.c b/mm/mempolicy.c index e0ae6997bbfb..f816b2ac5f52 100644 --- a/mm/mempolicy.c +++ b/mm/mempolicy.c @@ -1097,7 +1097,7 @@ static int migrate_to_node(struct mm_struct *mm, int source, int dest, * need migration. Between passing in the full user address * space range and MPOL_MF_DISCONTIG_OK, this call can not fail. */ - VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))); + VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL | MPOL_MF_LAZY))); queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask, flags | MPOL_MF_DISCONTIG_OK, &pagelist);
Some of our services are quite performance sensitive and actually NUMA-aware designed, aka numa-services. The SLOs can be easily violated when co-locate these services with other workloads. Thus they are granted to occupy the whole one or several NUMA nodes according to their quota. When a NUMA node is assigned to numa-service, the workload on that node needs to be moved away fast and complete. The main aspects we cared about on the eviction are as follows: a) it should complete soon enough so that numa-services won’t wait too long to hurt user experience b) the workloads to be evicted could have massive usage on memory, and migrating such amount of memory may lead to a sudden severe performance drop lasting tens of seconds that some certain workloads may not afford c) the impact of the eviction should be limited within the source and destination nodes d) cgroup interface is preferred So we come to a thought that: 1) fire up numa-services without waiting for memory migration 2) memory migration can be done asynchronously by using spare memory bandwidth AutoNUMA seems to be a solution, but its scope is global which violates c&d. And cpuset.memory_migrate performs in a synchronous fashion which breaks a&b. So a mixture of them, the new cgroup2 interface cpuset.mems.migration, is introduced. The new cpuset.mems.migration supports three modes: - "none" mode, meaning migration disabled - "sync" mode, which is exactly the same as the cgroup v1 interface cpuset.memory_migrate - "lazy" mode, when walking through all the pages, unlike cpuset.memory_migrate, it only sets pages to protnone, and numa faults triggered by later touch will handle the movement. See next patch for detailed information. Signed-off-by: Abel Wu <wuyun.abel@bytedance.com> --- kernel/cgroup/cpuset.c | 104 ++++++++++++++++++++++++++++++++--------- mm/mempolicy.c | 2 +- 2 files changed, 84 insertions(+), 22 deletions(-)