severely bloated slabs
12 hours after running dbench, tiobench, and a couple of others:
cache active alloc %util
buffer_head: 3781KB 132278KB 2.85
pte_chain: 46KB 520KB 8.98
dentry_cache: 8256KB 47662KB 17.32
vm_area_struct: 226KB 817KB 27.75
bio: 98KB 295KB 33.26
biovec-1: 25KB 75KB 33.45
size-32: 377KB 1098KB 34.38
size-64: 251KB 700KB 35.93
proc_inode_cache: 517KB 1157KB 44.74
task_struct: 788KB 1697KB 46.47
size-4096: 2480KB 5240KB 47.32
signal_act: 688KB 1379KB 49.86
skbuff_head_cache: 384KB 738KB 52.3
tcp_sock: 465KB 767KB 60.63
radix_tree_node: 9533KB 14851KB 64.18
files_cache: 487KB 677KB 71.96
sock_inode_cache: 248KB 341KB 72.74
sgpool-8: 195KB 255KB 76.83
size-2048: 1132KB 1372KB 82.50
size-256: 606KB 701KB 86.52
tcp_open_request: 54KB 62KB 87.43132MB of ZONE_NORMAL on a 16GB i386 box tied up in buffer_head slabs when all of 3% of it is in use gives me the willies. Periodic slab pruning anyone? Might be useful in addition to slab-in-lru.
Cheers,
Bill
--
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/
Quoted 3 lines
132MB of
ZONE_NORMALon a 16GB i386 box tied up inbuffer_headslabs when all of 3% of it is in use gives me the willies. Periodic slab pruning anyone? Might be useful in addition to slab-in-lru.
As long as we give this up under memory pressure, why does this matter?
M.
--
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/
Quoted 3 lines · At some point in the past, I wrote:
Quoted 3 lines
132MB of
ZONE_NORMALon a 16GB i386 box tied up inbuffer_headslabs when all of 3% of it is in use gives me the willies. Periodic slab pruning anyone? Might be useful in addition to slab-in-lru.
Quoted 2 lines · Martin J. Bligh
As long as we give this up under memory pressure, why does this matter?
Mostly a question of response time and long idle times being a good indicator of upcoming workload shifts. I'd say it's behaving as designed, but not as desired.
Cheers,
Bill
--
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/
Quoted 5 lines · William Lee Irwin III
Quoted 3 lines · At some point in the past, I wrote:
Quoted 3 lines
132MB of
ZONE_NORMALon a 16GB i386 box tied up inbuffer_headslabs when all of 3% of it is in use gives me the willies. Periodic slab pruning anyone? Might be useful in addition to slab-in-lru.
Quoted 7 lines
Quoted 2 lines · Martin J. Bligh
As long as we give this up under memory pressure, why does this matter?
Mostly a question of response time and long idle times being a good indicator of upcoming workload shifts. I'd say it's behaving as designed, but not as desired.
Was this with full rmap + slablru or with a linus kernel? With slablru I would expect this to happen to some extent. When vm pressure picks up slablru is fast to free the 'old' slabs... On the other hand, if periodic prunes are really a good idea, it would be easy to have slablru do them for us. As it stands now, slablru adds a flag bit to each slab cache telling slablru to prune the cache instead of just the page encounted. This flag gets set when we are able to add a page to the lru (when the pagemap_lru lock is busy). I would not be hard to set this flag under other conditions.
Ed
--
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/
Quoted 3 lines · William Lee Irwin III
Quoted 3 lines
Mostly a question of response time and long idle times being a good indicator of upcoming workload shifts. I'd say it's behaving as designed, but not as desired.
Quoted 8 lines · Ed Tomlinson
Was this with full rmap + slablru or with a linus kernel? With slablru I would expect this to happen to some extent. When vm pressure picks up slablru is fast to free the 'old' slabs... On the other hand, if periodic prunes are really a good idea, it would be easy to have slablru do them for us. As it stands now, slablru adds a flag bit to each slab cache telling slablru to prune the cache instead of just the page encounted. This flag gets set when we are able to add a page to the lru (when the
pagemap_lrulock is busy). I would not be hard to set this flag under other conditions.
Sounds reasonable, not sure what the general consensus is on periodic prunes, though. I'm mostly thinking of the desktop workload for this one, as servers are probably not going to really care. But I'm not a good testcase for this one as even my "desktop" usage patterns are atypical.
Cheers,
Bill
--
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/