It is tricky to quantify the basic cost of automatic NUMA placement in a
meaningful manner. This patch adds some vmstats that can be used as part
of a basic costing model.
u = basic unit = sizeof(void *)
Ca = cost of struct page access = sizeof(struct page) / u
Cpte = Cost PTE access = Ca
Cupdate = Cost PTE update = (2 * Cpte) + (2 * Wlock)
where Cpte is incurred twice for a read and a write and Wlock
is a constant representing the cost of taking or releasing a
lock
Cnumahint = Cost of a minor page fault = some high constant e.g. 1000
Cpagerw = Cost to read or write a full page = Ca + PAGE_SIZE/u
Ci = Cost of page isolation = Ca + Wi
where Wi is a constant that should reflect the approximate cost
of the locking operation
Cpagecopy = Cpagerw + (Cpagerw * Wnuma) + Ci + (Ci * Wnuma)
where Wnuma is the approximate NUMA factor. 1 is local. 1.2
would imply that remote accesses are 20% more expensive
Balancing cost = Cpte * numa_pte_updates +
Cnumahint * numa_hint_faults +
Ci * numa_pages_migrated +
Cpagecopy * numa_pages_migrated
Note that numa_pages_migrated is used as a measure of how many pages
were isolated even though it would miss pages that failed to migrate. A
vmstat counter could have been added for it but the isolation cost is
pretty marginal in comparison to the overall cost so it seemed overkill.
The ideal way to measure automatic placement benefit would be to count
the number of remote accesses versus local accesses and do something like
benefit = (remote_accesses_before - remove_access_after) * Wnuma
but the information is not readily available. As a workload converges, the
expection would be that the number of remote numa hints would reduce to 0.
convergence = numa_hint_faults_local / numa_hint_faults
where this is measured for the last N number of
numa hints recorded. When the workload is fully
converged the value is 1.
This can measure if the placement policy is converging and how fast it is
doing it.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Add a 1 second delay before starting to scan the working set of
a task and starting to balance it amongst nodes.
[ note that before the constant per task WSS sampling rate patch
the initial scan would happen much later still, in effect that
patch caused this regression. ]
The theory is that short-run tasks benefit very little from NUMA
placement: they come and go, and they better stick to the node
they were started on. As tasks mature and rebalance to other CPUs
and nodes, so does their NUMA placement have to change and so
does it start to matter more and more.
In practice this change fixes an observable kbuild regression:
# [ a perf stat --null --repeat 10 test of ten bzImage builds to /dev/shm ]
!NUMA:
45.291088843 seconds time elapsed ( +- 0.40% )
45.154231752 seconds time elapsed ( +- 0.36% )
+NUMA, no slow start:
46.172308123 seconds time elapsed ( +- 0.30% )
46.343168745 seconds time elapsed ( +- 0.25% )
+NUMA, 1 sec slow start:
45.224189155 seconds time elapsed ( +- 0.25% )
45.160866532 seconds time elapsed ( +- 0.17% )
and it also fixes an observable perf bench (hackbench) regression:
# perf stat --null --repeat 10 perf bench sched messaging
-NUMA:
-NUMA: 0.246225691 seconds time elapsed ( +- 1.31% )
+NUMA no slow start: 0.252620063 seconds time elapsed ( +- 1.13% )
+NUMA 1sec delay: 0.248076230 seconds time elapsed ( +- 1.35% )
The implementation is simple and straightforward, most of the patch
deals with adding the /proc/sys/kernel/numa_balancing_scan_delay_ms tunable
knob.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
[ Wrote the changelog, ran measurements, tuned the default. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Previously, to probe the working set of a task, we'd use
a very simple and crude method: mark all of its address
space PROT_NONE.
That method has various (obvious) disadvantages:
- it samples the working set at dissimilar rates,
giving some tasks a sampling quality advantage
over others.
- creates performance problems for tasks with very
large working sets
- over-samples processes with large address spaces but
which only very rarely execute
Improve that method by keeping a rotating offset into the
address space that marks the current position of the scan,
and advance it by a constant rate (in a CPU cycles execution
proportional manner). If the offset reaches the last mapped
address of the mm then it then it starts over at the first
address.
The per-task nature of the working set sampling functionality in this tree
allows such constant rate, per task, execution-weight proportional sampling
of the working set, with an adaptive sampling interval/frequency that
goes from once per 100ms up to just once per 8 seconds. The current
sampling volume is 256 MB per interval.
As tasks mature and converge their working set, so does the
sampling rate slow down to just a trickle, 256 MB per 8
seconds of CPU time executed.
This, beyond being adaptive, also rate-limits rarely
executing systems and does not over-sample on overloaded
systems.
[ In AutoNUMA speak, this patch deals with the effective sampling
rate of the 'hinting page fault'. AutoNUMA's scanning is
currently rate-limited, but it is also fundamentally
single-threaded, executing in the knuma_scand kernel thread,
so the limit in AutoNUMA is global and does not scale up with
the number of CPUs, nor does it scan tasks in an execution
proportional manner.
So the idea of rate-limiting the scanning was first implemented
in the AutoNUMA tree via a global rate limit. This patch goes
beyond that by implementing an execution rate proportional
working set sampling rate that is not implemented via a single
global scanning daemon. ]
[ Dan Carpenter pointed out a possible NULL pointer dereference in the
first version of this patch. ]
Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com>
Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
[ Wrote changelog and fixed bug. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
NOTE: This patch is based on "sched, numa, mm: Add fault driven
placement and migration policy" but as it throws away all the policy
to just leave a basic foundation I had to drop the signed-offs-by.
This patch creates a bare-bones method for setting PTEs pte_numa in the
context of the scheduler that when faulted later will be faulted onto the
node the CPU is running on. In itself this does nothing useful but any
placement policy will fundamentally depend on receiving hints on placement
from fault context and doing something intelligent about it.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
The use of MPOL_NOOP and MPOL_MF_LAZY to allow an application to
explicitly request lazy migration is a good idea but the actual
API has not been well reviewed and once released we have to support it.
For now this patch prevents an application using the services. This
will need to be revisited.
Signed-off-by: Mel Gorman <mgorman@suse.de>
This patch converts change_prot_numa() to use change_protection(). As
pte_numa and friends check the PTE bits directly it is necessary for
change_protection() to use pmd_mknuma(). Hence the required
modifications to change_protection() are a little clumsy but the
end result is that most of the numa page table helpers are just one or
two instructions.
Signed-off-by: Mel Gorman <mgorman@suse.de>
NOTE: Once again there is a lot of patch stealing and the end result
is sufficiently different that I had to drop the signed-offs.
Will re-add if the original authors are ok with that.
This patch adds another mbind() flag to request "lazy migration". The
flag, MPOL_MF_LAZY, modifies MPOL_MF_MOVE* such that the selected
pages are marked PROT_NONE. The pages will be migrated in the fault
path on "first touch", if the policy dictates at that time.
"Lazy Migration" will allow testing of migrate-on-fault via mbind().
Also allows applications to specify that only subsequently touched
pages be migrated to obey new policy, instead of all pages in range.
This can be useful for multi-threaded applications working on a
large shared data area that is initialized by an initial thread
resulting in all pages on one [or a few, if overflowed] nodes.
After PROT_NONE, the pages in regions assigned to the worker threads
will be automatically migrated local to the threads on 1st touch.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Note: Based on "mm/mpol: Use special PROT_NONE to migrate pages" but
sufficiently different that the signed-off-bys were dropped
Combine our previous _PAGE_NUMA, mpol_misplaced and migrate_misplaced_page()
pieces into an effective migrate on fault scheme.
Note that (on x86) we rely on PROT_NONE pages being !present and avoid
the TLB flush from try_to_unmap(TTU_MIGRATION). This greatly improves the
page-migration performance.
Based-on-work-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Mel Gorman <mgorman@suse.de>
If we have to avoid migrating to a node that is nearly full, put page
and return zero.
Signed-off-by: Hillf Danton <dhillf@gmail.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Note: This was originally based on Peter's patch "mm/migrate: Introduce
migrate_misplaced_page()" but borrows extremely heavily from Andrea's
"autonuma: memory follows CPU algorithm and task/mm_autonuma stats
collection". The end result is barely recognisable so signed-offs
had to be dropped. If original authors are ok with it, I'll
re-add the signed-off-bys.
Add migrate_misplaced_page() which deals with migrating pages from
faults.
Based-on-work-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Based-on-work-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Based-on-work-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
This patch provides a new function to test whether a page resides
on a node that is appropriate for the mempolicy for the vma and
address where the page is supposed to be mapped. This involves
looking up the node where the page belongs. So, the function
returns that node so that it may be used to allocated the page
without consulting the policy again.
A subsequent patch will call this function from the fault path.
Because of this, I don't want to go ahead and allocate the page, e.g.,
via alloc_page_vma() only to have to free it if it has the correct
policy. So, I just mimic the alloc_page_vma() node computation
logic--sort of.
Note: we could use this function to implement a MPOL_MF_STRICT
behavior when migrating pages to match mbind() mempolicy--e.g.,
to ensure that pages in an interleaved range are reinterleaved
rather than left where they are when they reside on any page in
the interleave nodemask.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
[ Added MPOL_F_LAZY to trigger migrate-on-fault;
simplified code now that we don't have to bother
with special crap for interleaved ]
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>
This patch augments the MPOL_MF_LAZY feature by adding a "NOOP" policy
to mbind(). When the NOOP policy is used with the 'MOVE and 'LAZY
flags, mbind() will map the pages PROT_NONE so that they will be
migrated on the next touch.
This allows an application to prepare for a new phase of operation
where different regions of shared storage will be assigned to
worker threads, w/o changing policy. Note that we could just use
"default" policy in this case. However, this also allows an
application to request that pages be migrated, only if necessary,
to follow any arbitrary policy that might currently apply to a
range of pages, without knowing the policy, or without specifying
multiple mbind()s for ranges with different policies.
[ Bug in early version of mpol_parse_str() reported by Fengguang Wu. ]
Bug-Reported-by: Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Note: This patch started as "mm/mpol: Create special PROT_NONE
infrastructure" and preserves the basic idea but steals *very*
heavily from "autonuma: numa hinting page faults entry points" for
the actual fault handlers without the migration parts. The end
result is barely recognisable as either patch so all Signed-off
and Reviewed-bys are dropped. If Peter, Ingo and Andrea are ok with
this version, I will re-add the signed-offs-by to reflect the history.
In order to facilitate a lazy -- fault driven -- migration of pages, create
a special transient PAGE_NUMA variant, we can then use the 'spurious'
protection faults to drive our migrations from.
The meaning of PAGE_NUMA depends on the architecture but on x86 it is
effectively PROT_NONE. Actual PROT_NONE mappings will not generate these
NUMA faults for the reason that the page fault code checks the permission on
the VMA (and will throw a segmentation fault on actual PROT_NONE mappings),
before it ever calls handle_mm_fault.
[dhillf@gmail.com: Fix typo]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
When we split a transparent hugepage, transfer the NUMA type from the
pmd to the pte if needed.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Introduce FOLL_NUMA to tell follow_page to check
pte/pmd_numa. get_user_pages must use FOLL_NUMA, and it's safe to do
so because it always invokes handle_mm_fault and retries the
follow_page later.
KVM secondary MMU page faults will trigger the NUMA hinting page
faults through gup_fast -> get_user_pages -> follow_page ->
handle_mm_fault.
Other follow_page callers like KSM should not use FOLL_NUMA, or they
would fail to get the pages if they use follow_page instead of
get_user_pages.
[ This patch was picked up from the AutoNUMA tree. ]
Originally-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
[ ported to this tree. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Implement pte_numa and pmd_numa.
We must atomically set the numa bit and clear the present bit to
define a pte_numa or pmd_numa.
Once a pte or pmd has been set as pte_numa or pmd_numa, the next time
a thread touches a virtual address in the corresponding virtual range,
a NUMA hinting page fault will trigger. The NUMA hinting page fault
will clear the NUMA bit and set the present bit again to resolve the
page fault.
The expectation is that a NUMA hinting page fault is used as part
of a placement policy that decides if a page should remain on the
current node or migrated to a different node.
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
The objective of _PAGE_NUMA is to be able to trigger NUMA hinting page
faults to identify the per NUMA node working set of the thread at
runtime.
Arming the NUMA hinting page fault mechanism works similarly to
setting up a mprotect(PROT_NONE) virtual range: the present bit is
cleared at the same time that _PAGE_NUMA is set, so when the fault
triggers we can identify it as a NUMA hinting page fault.
_PAGE_NUMA on x86 shares the same bit number of _PAGE_PROTNONE (but it
could also use a different bitflag, it's up to the architecture to
decide).
It would be confusing to call the "NUMA hinting page faults" as
"do_prot_none faults". They're different events and _PAGE_NUMA doesn't
alter the semantics of mprotect(PROT_NONE) in any way.
Sharing the same bitflag with _PAGE_PROTNONE in fact complicates
things: it requires us to ensure the code paths executed by
_PAGE_PROTNONE remains mutually exclusive to the code paths executed
by _PAGE_NUMA at all times, to avoid _PAGE_NUMA and _PAGE_PROTNONE to
step into each other toes.
Because we want to be able to set this bitflag in any established pte
or pmd (while clearing the present bit at the same time) without
losing information, this bitflag must never be set when the pte and
pmd are present, so the bitflag picked for _PAGE_NUMA usage, must not
be used by the swap entry format.
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Compaction already has tracepoints to count scanned and isolated pages
but it requires that ftrace be enabled and if that information has to be
written to disk then it can be disruptive. This patch adds vmstat counters
for compaction called compact_migrate_scanned, compact_free_scanned and
compact_isolated.
With these counters, it is possible to define a basic cost model for
compaction. This approximates of how much work compaction is doing and can
be compared that with an oprofile showing TLB misses and see if the cost of
compaction is being offset by THP for example. Minimally a compaction patch
can be evaluated in terms of whether it increases or decreases cost. The
basic cost model looks like this
Fundamental unit u: a word sizeof(void *)
Ca = cost of struct page access = sizeof(struct page) / u
Cmc = Cost migrate page copy = (Ca + PAGE_SIZE/u) * 2
Cmf = Cost migrate failure = Ca * 2
Ci = Cost page isolation = (Ca + Wi)
where Wi is a constant that should reflect the approximate
cost of the locking operation.
Csm = Cost migrate scanning = Ca
Csf = Cost free scanning = Ca
Overall cost = (Csm * compact_migrate_scanned) +
(Csf * compact_free_scanned) +
(Ci * compact_isolated) +
(Cmc * pgmigrate_success) +
(Cmf * pgmigrate_failed)
Where the values are read from /proc/vmstat.
This is very basic and ignores certain costs such as the allocation cost
to do a migrate page copy but any improvement to the model would still
use the same vmstat counters.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
The pgmigrate_success and pgmigrate_fail vmstat counters tells the user
about migration activity but not the type or the reason. This patch adds
a tracepoint to identify the type of page migration and why the page is
being migrated.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
The compact_pages_moved and compact_pagemigrate_failed events are
convenient for determining if compaction is active and to what
degree migration is succeeding but it's at the wrong level. Other
users of migration may also want to know if migration is working
properly and this will be particularly true for any automated
NUMA migration. This patch moves the counters down to migration
with the new events called pgmigrate_success and pgmigrate_fail.
The compact_blocks_moved counter is removed because while it was
useful for debugging initially, it's worthless now as no meaningful
conclusions can be drawn from its value.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reuse the NUMA code's 'modified page protections' count that
change_protection() computes and skip the TLB flush if there's
no changes to a range that sys_mprotect() modifies.
Given that mprotect() already optimizes the same-flags case
I expected this optimization to dominantly trigger on
CONFIG_NUMA_BALANCING=y kernels - but even with that feature
disabled it triggers rather often.
There's two reasons for that:
1)
While sys_mprotect() already optimizes the same-flag case:
if (newflags == oldflags) {
*pprev = vma;
return 0;
}
and this test works in many cases, but it is too sharp in some
others, where it differentiates between protection values that the
underlying PTE format makes no distinction about, such as
PROT_EXEC == PROT_READ on x86.
2)
Even where the pte format over vma flag changes necessiates a
modification of the pagetables, there might be no pagetables
yet to modify: they might not be instantiated yet.
During a regular desktop bootup this optimization hits a couple
of hundred times. During a Java test I measured thousands of
hits.
So this optimization improves sys_mprotect() in general, not just
CONFIG_NUMA_BALANCING=y kernels.
[ We could further increase the efficiency of this optimization if
change_pte_range() and change_huge_pmd() was a bit smarter about
recognizing exact-same-value protection masks - when the hardware
can do that safely. This would probably further speed up mprotect(). ]
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
With transparent hugepage support, handle_mm_fault() has to be careful
that a normal PMD has been established before handling a PTE fault. To
achieve this, it used __pte_alloc() directly instead of pte_alloc_map
as pte_alloc_map is unsafe to run against a huge PMD. pte_offset_map()
is called once it is known the PMD is safe.
pte_alloc_map() is smart enough to check if a PTE is already present
before calling __pte_alloc but this check was lost. As a consequence,
PTEs may be allocated unnecessarily and the page table lock taken.
Thi useless PTE does get cleaned up but it's a performance hit which
is visible in page_test from aim9.
This patch simply re-adds the check normally done by pte_alloc_map to
check if the PTE needs to be allocated before taking the page table
lock. The effect is noticable in page_test from aim9.
AIM9
2.6.38-vanilla 2.6.38-checkptenone
creat-clo 446.10 ( 0.00%) 424.47 (-5.10%)
page_test 38.10 ( 0.00%) 42.04 ( 9.37%)
brk_test 52.45 ( 0.00%) 51.57 (-1.71%)
exec_test 382.00 ( 0.00%) 456.90 (16.39%)
fork_test 60.11 ( 0.00%) 67.79 (11.34%)
MMTests Statistics: duration
Total Elapsed Time (seconds) 611.90 612.22
(While this affects 2.6.38, it is a performance rather than a
functional bug and normally outside the rules -stable. While the big
performance differences are to a microbench, the difference in fork
and exec performance may be significant enough that -stable wants to
consider the patch)
Reported-by: Raz Ben Yehuda <raziebe@gmail.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Rik van Riel <riel@redhat.com>
[ Picked this up from the AutoNUMA tree to help
it upstream and to allow apples-to-apples
performance comparisons. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Mel Gorman
Peter Zijlstra
Lee Schermerhorn
Andrea Arcangeli
Ingo Molnar