memory page alloc minor cleanups
Subsystems (6)
RFC 1/1 memory page alloc minor cleanups
While coding up various alternative performance improvements to the zonelist scanning below __alloc_pages(), I tripped over a few minor code style and layout nits in mm/page_alloc.c
I noticed that Nick had a couple of these same nits in one of his patches - so I hesitate to push this patch without sync'ing with him, to minimize conflicts over more important patches.
The removal of the NULL zone check needs approval by someone who knows this code better than I do -- I could have broken something with this change.
Changes include:
1) s/freeliest/freelist/ spelling fix
2) Check for NULL *z zone seems useless - even if it could happen, so what? Perhaps we should have a check later on if we are faced with an allocation request that is not allowed to fail - shouldn't that be a serious kernel error, passing an empty zonelist with a mandate to not fail?
3) Initializing 'z' to zonelist->zones can wait until after the first get_page_from_freelist() fails; we only use 'z' in the wakeup_kswapd() loop, so let's initialize 'z' there, in a 'for' loop. Seems clearer.
4) Remove superfluous braces around a break
5) Fix a couple errant spaces
6) Adjust indentation on the cpuset_zone_allowed() check, to match the lines just before it -- seems easier to read in this case.
7) Add another set of braces to the zone_watermark_ok logic
Changes (4) and (7) I stole from some patch of Nick's.
--- mm/page_alloc.c | 27 ++++++++++----------------- 1 file changed, 10 insertions(+), 17 deletions(-)
mm/page_alloc.c +10 -17
@@ -497,7 +497,7 @@ static void free_one_page(struct zone *zspin_lock (&zone->lock); zone->all_unreclaimable =0 ; zone->pages_scanned =0 ; -__free_one_page (page, zone ,order); +__free_one_page (page, zone, order);spin_unlock (&zone->lock); }
@@ -937,7 +937,7 @@ int zone_watermark_ok(struct zone *z, in }/* -* get_page_from_freeliest goes through the zonelist trying to allocate +* get_page_from_freelist goes through the zonelist trying to allocate * a page. */ static struct page *
@@ -959,8 +959,8 @@ get_page_from_freelist(gfp_t gfp_mask, u zone->zone_pgdat != zonelist->zones[0]->zone_pgdat))break ;if ((alloc_flags &ALLOC_CPUSET ) && - !cpuset_zone_allowed (zone, gfp_mask)) -continue ; + !cpuset_zone_allowed (zone, gfp_mask)) +continue ;if (!(alloc_flags &ALLOC_NO_WATERMARKS )) {unsigned long mark;
@@ -970,17 +970,18 @@ get_page_from_freelist(gfp_t gfp_mask, u mark = zone->pages_low; else mark = zone->pages_high; -if (!zone_watermark_ok (zone , order, mark, - classzone_idx, alloc_flags)) +if (!zone_watermark_ok (zone, order, mark, + classzone_idx, alloc_flags)) {if (!zone_reclaim_mode || !zone_reclaim (zone, gfp_mask, order))continue ; + } } page =buffered_rmqueue (zonelist, zone, order, gfp_mask); -if (page) { +if (page)break ; - } + }while (*(++z) !=NULL );return page; }
@@ -1056,21 +1057,13 @@ __alloc_pages(gfp_t gfp_mask, unsigned imight_sleep_if (wait); restart: - z = zonelist->zones;/* the list of zones suitable for gfp_mask */ - -if (unlikely (*z ==NULL )) { -/* Should this ever happen?? */ -return NULL ; - } - page =get_page_from_freelist (gfp_mask|__GFP_HARDWALL, order, zonelist,ALLOC_WMARK_LOW |ALLOC_CPUSET );if (page) goto got_pg; -do { +for (z = zonelist->zones; *z; z++)wakeup_kswapd (*z, order); - }while (*(++z));/* * OK, we're below the kswapd watermark and have kicked background
Signature · 9 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.650.933.1373
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>RFC 1/1 memory page_alloc zonelist caching speedup
Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them.
Remembers the zones in a zonelist that were short of free memory in the last second. And it stashes a zone-to-node table in the zonelist struct, to optimize that conversion (minimize its cache footprint.)
It applies to 2.6.18-mm3, plus the patch:
memory page alloc minor cleanups
I'll backport it to 2.6.18 so that Rohit can test it, in a subsequent post.
Recent changes:
This differs in a significant way from a similar patch that I posted a week ago. Now, instead of having a nodemask_t of recently full nodes, I have a bitmask of recently full zones. This solves a problem that last weeks patch had, which on systems with multiple zones per node (such as DMA zone) would take seeing any of these zones full as meaning that all zones on that node were full.
Also I changed names - from "zonelist faster" to "zonelist cache", as that seemed to better convey what we're doing here - caching some of the key zonelist state (for faster access.)
See below for some performance benchmark results. After all that discussion with David on why I didn't need them, I went and got some ;). I wanted to verify that I had not hurt the normal case of memory allocation noticeably. At least for my one little microbenchmark, I found (1) the normal case wasn't affected, and
(2) workloads that forced scanning across multiple nodes for memory improved up to 10% fewer System CPU cycles and lower elapsed clock time ('sys' and 'real'). Good. See details, below.
I didn't have the logic in get_page_from_freelist() for various full nodes and zone reclaim failures correct. That should be fixed up now - notice the new goto labels zonelist_scan, this_zone_full, and try_next_zone, in get_page_from_freelist().
There are two reasons I persued this alternative, over some earlier proposals that would have focused on optimizing the fake numa emulation case by caching the last useful zone:
1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems) have seen real customer loads where the cost to scan the zonelist was a problem, due to many nodes being full of memory before we got to a node we could use. Or at least, I think we have. This was related to me by another engineer, based on experiences from some time past. So this is not guaranteed. Most likely, though.
The following approach should help such real numa systems just as much as it helps fake numa systems, or any combination thereof.
2) The effort to distinguish fake from real numa, using node_distance, so that we could cache a fake numa node and optimize choosing it over equivalent distance fake nodes, while continuing to properly scan all real nodes in distance order, was going to require a nasty blob of zonelist and node distance munging.
The following approach has no new dependency on node distances or zone sorting.
See comment in the patch below for a description of what it actually does.
Technical details of note (or controversy):
- See the use of "zlc_active" and "did_zlc_setup" below, to delay adding any work for this new mechanism until we've looked at the first zone in zonelist. I figured the odds of the first zone having the memory we needed were high enough that we should just look there, first, then get fancy only if we need to keep looking.
- Some odd hackery was needed to add items to struct zonelist, while not tripping up the custom zonelists built by the mm/mempolicy.c code for MPOL_BIND. My usual wordy comments below explain this. Search for "MPOL_BIND".
- Some per-node data in the struct zonelist is now modified frequently, with no locking. Multiple CPU cores on a node could hit and mangle this data. The theory is that this is just performance hint data, and the memory allocator will work just fine despite any such mangling. The fields at risk are the struct 'zonelist_cache' fields 'fullzones' (a bitmask) and 'last_full_zap' (unsigned long jiffies). It should all be self correcting after at most a one second delay.
- This still does a linear scan of the same lengths as before. All I've optimized is making the scan faster, not algorithmically shorter. It is now able to scan a compact array of 'unsigned short' in the case of many full nodes, so one cache line should cover quite a few nodes, rather than each node hitting another one or two new and distinct cache lines.
- If both Andi and Nick don't find this too complicated, I will be (pleasantly) flabbergasted.
- I removed the comment claiming we only use one cachline's worth of zonelist. We seem, at least in the fake numa case, to have put the lie to that claim.
- I pay no attention to the various watermarks and such in this performance hint. A node could be marked full for one watermark, and then skipped over when searching for a page using a different watermark. I think that's actually quite ok, as it will tend to slightly increase the spreading of memory over other nodes, away from a memory stressed node.
===============
Performance - some benchmark results and analysis:
This benchmark runs a memory hog program that uses multiple threads to touch alot of memory as quickly as it can.
Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of the total 96 GBytes on the system, and using 1, 19, 37, or 55 threads (on a 56 CPU system.) System, user and real (elapsed) timings were recorded for each run, shown in units of seconds, in the table below.
Two kernels were tested - 2.6.18-mm3 and the same kernel with this zonelist caching patch added. The table also shows the difference in timings, between these two kernels, and what percentage improvement the zonelist caching sys time is over (lower than) the stock *-mm kernel.
number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent
GBs N ------------ -------------- ---------------- systime
mem threads sys user real sys user real sys user real better
12 1 153 24 177 151 24 176 -2 0 -1 1%
12 19 99 22 8 99 22 8 0 0 0 0% 12 37 111 25 6 112 25 6 1 0 0 -0% 12 55 115 25 5 110 23 5 -5 -2 0 4% 38 1 502 74 576 497 73 570 -5 -1 -6 0% 38 19 426 78 48 373 76 39 -53 -2 -9 12% 38 37 544 83 36 547 82 36 3 -1 0 -0% 38 55 501 77 23 511 80 24 10 3 1 -1% 64 1 917 125 1042 890 124 1014 -27 -1 -28 2% 64 19 1118 138 119 965 141 103 -153 3 -16 13% 64 37 1202 151 94 1136 150 81 -66 -1 -13 5% 64 55 1118 141 61 1072 140 58 -46 -1 -3 4% 90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4% 90 19 2392 199 192 2116 189 176 -276 -10 -16 11% 90 37 3313 238 175 2972 225 145 -341 -13 -30 10% 90 55 1948 210 104 1843 213 100 -105 3 -4 5%
Notes:
1) This test ran a memory hog program that started a specified number N of threads, and had each thread allocate and touch 1/N'th of the total memory to be used in the test run in a single loop, writing a constant word to memory, one store every 4096 bytes. Watching this test during some earlier trial runs, I would see each of these threads sit down on one CPU and stay there, for the remainder of the pass, a different CPU for each thread.
2) The 'real' column is not comparable to the 'sys' or 'user' columns. The 'real' column is seconds wall clock time elapsed, from beginning to end of that test pass. The 'sys' and 'user' columns are total CPU seconds spent on that test pass. For a 19 thread test run, for example, the sum of 'sys' and 'user' could be up to 19 times the number of 'real' elapsed wall clock seconds.
3) Tests were run on a fresh, single-user boot, to minimize the amount of memory already in use at the start of the test, and to minimize the amount of background activity that might interfere.
4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM.
5) Notice that the 'real' time gets large for the single thread runs, even though the measured 'sys' and 'user' times are modest. I'm not sure what that means - probably something to do with it being slow for one thread to be accessing memory along ways away. Perhaps the fake numa system, running ostensibly the same workload, would not show this substantial degradation of 'real' time for one thread on many nodes -- lets hope not.
6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs) ran quite efficiently, as one might expect. Each pair of threads needed to allocate and touch the memory on the node the two threads shared, a pleasantly parallizable workload.
7) The intermediate thread count passes, when asking for alot of memory forcing them to go to a few neighboring nodes, improved the most with this zonelist caching patch.
Conclusions:
* This zonelist cache patch probably makes little difference one way or the other for most workloads on real numa hardware, if those workloads avoid heavy off node allocations.
* For memory intensive workloads requiring substantial off-node allocations on real numa hardware, this patch improves both kernel and elapsed timings up to ten per-cent.
* For fake numa systems, I'm optimistic, but will have to leave that up to Rohit Seth to actually test (once I get him a 2.6.18 backport.)
===============
--- include/linux/cpuset.h | 2 include/linux/mmzone.h | 83 ++++++++++++++++++++- mm/mempolicy.c | 2 mm/page_alloc.c | 188 +++++++++++++++++++++++++++++++++++++++++++++++-- 4 files changed, 263 insertions(+), 12 deletions(-)
include/linux/mmzone.h +78 -5
@@ -315,19 +315,92 @@ struct zone {*/ #define DEF_PRIORITY 12 +#ifdef CONFIG_NUMA +/* +* We cache key information from each zonelist for smaller cache +* footprint when scanning for free pages in get_page_from_freelist(). +* +* 1) The BITMAP fullzones tracks which zones in a zonelist have come +* up short of free memory since the last time (last_fullzone_zap) +* we zero'd fullzones. +* 2) The array z_to_n[] maps each zone in the zonelist to its node +* id, so that we can efficiently evaluate whether that node is +* set in the current tasks mems_allowed. +* +* Both fullzones and z_to_n[] are one-to-one with the zonelist, +* indexed by a zones offset in the zonelist zones[] array. +* +* The get_page_from_freelist() routine does two scans. During the +* first scan, we skip zones whose corresponding bit in 'fullzones' +* is set or whose corresponding node in current->mems_allowed (which +* comes from cpusets) is not set. During the second scan, we bypass +* this zonelist_cache, to ensure we look methodically at each zone. +* +* Once per second, we zero out (zap) fullzones, forcing us to +* reconsider nodes that might have regained more free memory. +* The field last_full_zap is the time we last zapped fullzones. +* +* This mechanism reduces the amount of time we waste repeatedly +* reexaming zones for free memory when they just came up low on +* memory momentarilly ago. +* +* The zonelist_cache struct members logically belong in struct +* zonelist. However, the mempolicy zonelists constructed for +* MPOL_BIND are intentionally variable length (and usually much +* shorter). A general purpose mechanism for handling structs with +* multiple variable length members is more mechanism than we want +* here. We resort to some special case hackery instead. +* +* The MPOL_BIND zonelists don't need this zonelist_cache (in good +* part because they are shorter), so we put the fixed length stuff +* at the front of the zonelist struct, ending in a variable length +* zones[], as is needed by MPOL_BIND. +* +* Then we put the optional zonelist cache on the end of the zonelist +* struct. This optional stuff is found by a 'zlcache_ptr' pointer in +* the fixed length portion at the front of the struct. This pointer +* both enables us to find the zonelist cache, and in the case of +* MPOL_BIND zonelists, (which will just set the zlcache_ptr to NULL) +* to know that the zonelist cache is not there. +* +* The end result is that struct zonelists come in two flavors: +* 1) The full, fixed length version, shown below, and +* 2) The custom zonelists for MPOL_BIND. +* The custom MPOL_BIND zonelists have a NULL zlcache_ptr and no zlcache. +* +* Even though there may be multiple CPU cores on a node modifying +* fullzones or last_full_zap in the same zonelist_cache at the same +* time, we don't lock it. This is just hint data - if it is wrong now +* and then, the allocator will still function, perhaps a bit slower. +*/ + +#define MAX_ZONES_PER_ZONELIST (MAX_NUMNODES * MAX_NR_ZONES) + +struct zonelist_cache { + DECLARE_BITMAP(fullzones, MAX_ZONES_PER_ZONELIST); /* zone full? */ +unsigned short z_to_n[MAX_ZONES_PER_ZONELIST ];/* zone->nid */ +unsigned long last_full_zap;/* when last zap'd (jiffies) */ +}; +#else +struct zonelist_cache ; +#endif +/* * One allocation request operates on a zonelist. A zonelist * is a list of zones, the first one is the 'goal' of the * allocation, the other zones are fallback zones, in decreasing * priority. * -* Right now a zonelist takes up less than a cacheline. We never -* modify it apart from boot-up, and only a few indices are used, -* so despite the zonelist table being relatively big, the cache -* footprint of this construct is very small. +* If zlcache_ptr is not NULL, then it is just the address of zlcache, +* as explained above. If zlcache_ptr is NULL, there is no zlcache. */ +struct zonelist { -struct zone *zones[MAX_NUMNODES *MAX_NR_ZONES +1 ];// NULL delimited +struct zonelist_cache *zlcache_ptr;// NULL or &zlcache +struct zone *zones[MAX_ZONES_PER_ZONELIST +1 ];// NULL delimited +#ifdef CONFIG_NUMA +struct zonelist_cache zlcache;// optional ... +#endif };#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
mm/page_alloc.c +181 -7
245 lines not on this page · raw message
mm/mempolicy.c +2 -0
@@ -141,9 +141,11 @@ static struct zonelist *bind_zonelist(noenum zone_type k; max =1 +MAX_NR_ZONES *nodes_weight (*nodes); + max++;/* space for zlcache_ptr (see mmzone.h) */ zl =kmalloc (sizeof (struct zone *) * max,GFP_KERNEL );if (!zl)return NULL ; + zl->zlcache_ptr =NULL ; num =0 ;/* First put in the highest zones from all nodes, then all the next lower zones etc. Avoid empty zones because the memory allocator
include/linux/cpuset.h +2 -0
@@ -23,6 +23,7 @@ extern void cpuset_fork(struct task_struextern void cpuset_exit (struct task_struct *p);extern cpumask_t cpuset_cpus_allowed (struct task_struct *p);extern nodemask_t cpuset_mems_allowed (struct task_struct *p); +#define cpuset_current_mems_allowed (current->mems_allowed)void cpuset_init_current_mems_allowed (void );void cpuset_update_task_memory_state (void );#define cpuset_nodes_subset_current_mems_allowed (nodes) \
@@ -83,6 +84,7 @@ static inline nodemask_t cpuset_mems_allreturn node_possible_map; } +#define cpuset_current_mems_allowed (node_online_map)static inline void cpuset_init_current_mems_allowed (void ) {}static inline void cpuset_update_task_memory_state (void ) {}#define cpuset_nodes_subset_current_mems_allowed (nodes) (1)
Signature · 9 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.650.933.1373
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>Re: [RFC] memory page_alloc zonelist caching speedup
Quoted 3 lines · On Mon, 09 Oct 2006 03:54:57 -0700 Paul Jackson <[email protected]> wrote:
Optimize the critical zonelist scanning for free pages in the kernel memory allocator by caching the zones that were found to be full recently, and skipping them.
This doesn't exactly simplify the kernel, but the benchmark numbers are nice.
I worry about the one-second-expiry thing. Wall time is a pretty meaningless thing in the context of the page allocator and it doesn't seem appropriate to use it. A more appropriate measure of "time" in this context would be number-of-pages-allocated.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Re: [RFC] memory page_alloc zonelist caching speedup
Quoted 4 lines · Andrew
I worry about the one-second-expiry thing. Wall time is a pretty meaningless thing in the context of the page allocator and it doesn't seem appropriate to use it. A more appropriate measure of "time" in this context would be number-of-pages-allocated.
Yeah, maybe ...
Let's take a couple of extreme examples.
1) Let's say a compute intensive app is growing slowly, one page every few seconds. Do we really care whether or not we take the fast path or the slow path through the page allocator in this case? I doubt it.
Though, if it's been a while, I'd sooner take the slow path code and get the page placed exactly on the first zone that can provide it. Just because a cache hasn't been used in several seconds doesn't make it still useful. Maybe something we didn't count changed.
For this reason, I still think a time based expiration is useful.
2) Let's say you just got a sample petahertz processor from your favorite CPU vendor, with 64 cores and 4 TBytes of 10 picosecond RAM, all in one package. You can built, boot and test your entire distro in 4.2 seconds. The average teenager can do it in 2.7 seconds, because they have faster fingers. Life is good.
Yeah - well - in that case only resetting this cache 4 times in that entire build, boot and test cycle is retarded.
So that suggests we need two triggers on the cache expiration:
* the current time trigger, and
* a counter trigger - say every 1000 allocations.
Once either the count or the time trigger is hit, reset the cache.
I guess this means we add a counter to the zonelist_cache struct. Increment it each time we try to allocate a page from that zonelist. Trigger a zap (cache expiration) if the counter hits 1000, and clear the counter when we do the zap.
If we have many CPUs banging on one poor zonelist, this counter risks creating a warm cache line. Though since this is per zone (tends to be per node) and since the rest of this line of memory is stone cold, this is not likely to be a serious problem.
Normally, if I'm not sure I need a line of code, I don't code it. But if it makes others happier to have the extra code, and it seems harmless enough, then what the heck - add it.
Guess I should code up such a counter, so we can see how it looks.
I still doubt it matters ...
Signature · 9 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>Re: [RFC] memory page_alloc zonelist caching speedup
Quoted 4 lines · pj
I guess this means we add a counter to the
zonelist_cachestruct. Increment it each time we try to allocate a page from that zonelist. Trigger a zap (cache expiration) if the counter hits 1000, and clear the counter when we do the zap.
No, dang it. Not count allocs. Count frees.
My zonelist caching adapts immediately to another node being filled up due to allocs, by turning on another bit in the fullzones bitmask.
But it doesn't adapt immediately to memory coming free.
An application could say free up a big chunk of memory on its local node - perhaps by dropping the last reference to an anonymous memory region.
It would then reasonably expect that new allocations would come from the local node - right then - not starting some later time up to one second in the future.
I think I'd need a per-node counter of frees, incremented on each free, and checked by the zonelist caching to see if it should consider that node no longer full.
-However- that forces a per-node reference in the zonelist caching code as part of the scan for a free page. That is exactly what we were trying to avoid!
No. Not count frees either. Don't count anything.
I do not see how to count anything related to allocs or frees and then use that counter to throttle the zonelist caching, without re-introducing the lousy cache line footprint that I just got done shrinking.
That's why I like time based throttles. They are cheap. Dirt cheap. Infinitely scalable. And stupid as a pet rock ;).
P.S. I don't think that the above application, expecting instant reuse of node local memory, even though it had just been pushing allocations off-node, is a real problem. Anyone care to claim otherwise?
Signature · 9 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>Re: [RFC] memory page_alloc zonelist caching speedup
Quoted 6 lines · Paul Jackson
-However- that forces a per-node reference in the zonelist caching code as part of the scan for a free page. That is exactly what we were trying to avoid!
No. Not count frees either. Don't count anything.
When a free occurs for a given zone, increment its counter. If that reaches some threshold, zap that node in the nodemask so it's checked on the next alloc. All the infrastructure is already there for this support in your patch.
[ Note: [email protected] is no longer valid so I've removed it (again)
from the Cc list. My email address is [email protected]. ]
David
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Re: [RFC] memory page_alloc zonelist caching speedup
Quoted 4 lines
When a free occurs for a given zone, increment its counter. If that reaches some threshold, zap that node in the nodemask so it's checked on the next alloc. All the infrastructure is already there for this support in your patch.
It's not an issue of infrastructure. As you say, that's likely already there.
It's the inherent problem in scaling an N-by-N information flow, with tasks running on each of N nodes wanting to know the latest free counters on each of N nodes. This cannot be done with a small constant (or linear, but so small it is nearly constant) cache footprint for both the freers and allocators, avoiding hot cache lines.
In your phrasing, this shows up in the "zap that node in the nodemask" step.
We don't have -a- nodemask.
My latest patch has a bitmask (of length longer than a nodemask, typically) in each zonelist. No way do we want to walk down each zonelist, one each per node, per ZONE type, examining each zone to see if it's on our node of interest, so we can clear the corresponding bit in the bitmask. Not on every page free. Way too expensive.
Signature · 9 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>Re: [RFC] memory page_alloc zonelist caching speedup
Could it be worth to investigate more radical ideas? This gets way too complicated for me. Maybe drop the whole zone list generation idea and iterate over nodes in another way?
1. Have an allocator that is not node aware and can just deal with memory in up to 3 different zones. No NUMA at all.
2. Have another NUMA allocator that uses the node unaware allocator but implements its own way of handling the NUMA situation with proper fallbacks etc etc. Maybe we could then merge the allocation logic from mempolicy.c into the page allocator?
If 2 is generic enough then it can be used for other allocators as well (like slab, hugepages, uncaches allocators) and provide a coherent NUMA allocation handling for all allocators on NUMA.
It would be great if we could simplify and modularize the page allocator.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Re: [RFC] memory page_alloc zonelist caching speedup
Quoted 3 lines · Christoph
Could it be worth to investigate more radical ideas? This gets way too complicated for me. Maybe drop the whole zone list generation idea and iterate over nodes in another way?
Worth some thought.
I'll be surprised if this eliminates the usefulness of something like this zonelist caching patch, however.
Sooner or later, regardless of what shape data structures we have, we end up having to examine a bunch of nodes when allocating for workloads or numa emulated configurations that make heavy use of off-node allocations.
And when that happens, we end up with an N-squared information flow problem, needing to get information or at least hints as to which nodes have free pages to the tasks trying to allocate those pages.
But we really would rather not pay the price of even a linear scan over N nodes, in either the tasks freeing pages, nor in the tasks allocating them.
The best I've been able to do, in this patch, is:
1) compact the information, to minimize the cache line footprint, and
2) have the allocators get by on incomplete information, essentially doing the first scan based on remembering which nodes were recently noticed to be full.
I predict that regardless of the shape (zonelists, nodemasks or whatever) of the placement information coming into the core routine of our allocator, we will still need some sort of caching like this, bolted onto the side, for the cases making heavy use of off-node allocations.
So I would not use disgust at the added complexity of this zonelist caching patch to justify changing the fundamental zonelist structures used to drive the kernel allocator.
Signature · 9 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>Re: [RFC] memory page_alloc zonelist caching speedup
In reply to Paul JacksonQuoted 4 lines · pj
2) Let's say you just got a sample petahertz processor from your favorite CPU vendor, with 64 cores and 4 TBytes of 10 picosecond RAM, all in one package. You can built, boot and test your entire distro in 4.2 seconds.
This silly example motivates changing my one second (1 * HZ) constant timeout on the zonelist cache to a variable, computed at boottime as a simple minded function of bogomips or clock speed or some such.
At the (slow) rate of CPU frequency increases the last few years, we've got a while before we need to worry about this change.
Signature · 9 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Paul Jackson
Acked-by Christoph LameterThat's odd. The copy of Christoph's Ack that I got directly had the one line body:
but the copy that I got via the linux-mm email list just had the standard linux-mm email list footer in its body, and not the above Acked line from Christoph.
Something in the path this message took through linux-mm stripped off Christoph's Acked-by line.
Signature · 9 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>Quoted 11 lines · Paul Jackson
That's odd. The copy of Christoph's Ack that I got directly had the one line body:
but the copy that I got via the linux-mm email list just had the standard linux-mm email list footer in its body, and not the above Acked line from Christoph.
Something in the path this message took through linux-mm stripped off Christoph's Acked-by line.
Yes we had this before. One needs to add some text or a blank line so that linux-mm does not eat it.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 9 lines · Paul Jackson
From: Paul Jackson <[email protected]>
While coding up various alternative performance improvements to the zonelist scanning below
__alloc_pages(), I tripped over a few minor code style and layout nits inmm/page_alloc.cI noticed that Nick had a couple of these same nits in one of his patches - so I hesitate to push this patch without sync'ing with him, to minimize conflicts over more important patches.
Ah, syncing up won't be difficult.
Quoted 5 lines
The removal of the NULL zone check needs approval by someone who knows this code better than I do -- I could have broken something with this change.
Quoted 7 lines
Changes include:
1) s/freeliest/freelist/ spelling fix
2) Check for NULL *z zone seems useless - even if it could happen, so what? Perhaps we should have a check later on if we are faced with an allocation request that is not allowed to fail - shouldn't that be a serious kernel error, passing an empty zonelist with a mandate to not fail?
Would it be better to ensure an empty zonelist is never passed down?
Otherwise, it's fine by me.
Signature · 8 lines
SUSE Labs, Novell Inc. Send instant messages to your online friends http://au.messenger.yahoo.com -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Probably in response to my patch lines:
@@ -1056,21 +1057,13 @@ __alloc_pages(gfp_t gfp_mask, unsigned i
...
- if (unlikely(*z == NULL)) {
- /* Should this ever happen?? */
- return NULL;
- }
Quoted 1 line · Nick
Would it be better to ensure an empty zonelist is never passed down?
Are you saying we should leave this empty zonelist check where it was, or we should somehow ensure that we never get to __alloc_pages with an empty zonelist in the first place? Not clear ...
What seems clear to me is that this check is in the wrong place, and if needed, is the wrong check.
The check is not needed right there. If we have an empty zonelist, then that just makes the zonelist scanning go all the faster ;). Harmless, silly, but rare.
Not until much deeper in the allocation code, when we have to make some hard choices, like oom or panic or loop forever (hopelessly) looking for pages off an empty zonelist, do we actually have to worry about empty zonelists.
So either:
* the check is not needed, if empty zonelists can't happen, or
* the check should be moved out of the hot spot it is in now, where it has no need of being, to where it is needed, lower down, in less frequently executed code.
And if it is needed, the logic of the check seems slightly oversimplified:
I'd think it should consider (1) allocations requests that can fail in which case we return NULL, separately from (2) allocation requests that cannot fail in which case we are on an impossible mission, as the caller is insisting that we do not fail to find a page on an empty list.
Perhaps in this second case, we pick the local nodes default full sized zonelist and find a page for our demanding caller that way.
Signature · 9 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>Quoted 3 lines · pj
The check is not needed right there. If we have an empty zonelist, then that just makes the zonelist scanning go all the faster ;). Harmless, silly, but rare.
I should read the code before spouting off ... ;).
The get_page_from_freelist() code assumes in many places that there is at least one zone in the zonelist. It will barf all over the place if zonelist->zones[0] is not a valid pointer.
Either this check for an empty zonelist at the top of __alloc_pages() stays, or it becomes some kind of BUG() or someone more confident than I removes it.
I'll be sending a patch to restore that check for an empty zonelist, shortly.
Signature · 9 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>