net: Optimize snmp stat aggregation by walking all the percpu data at once
authorRaghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Sun, 30 Aug 2015 05:59:42 +0000 (11:29 +0530)
committerDavid S. Miller <davem@davemloft.net>
Mon, 31 Aug 2015 04:48:58 +0000 (21:48 -0700)
commita3a773726c9f9ba2e87fd8ad8e36feff5f6ffd8e
treebc03244dc5119515d96c6c9a64b0c871ad554385
parentc4c6bc314618f60ba69b0cbf93e506e4c38a11d2
net: Optimize snmp stat aggregation by walking all the percpu data at once

Docker container creation linearly increased from around 1.6 sec to 7.5 sec
(at 1000 containers) and perf data showed 50% ovehead in snmp_fold_field.

reason: currently __snmp6_fill_stats64 calls snmp_fold_field that walks
through per cpu data of an item (iteratively for around 36 items).

idea: This patch tries to aggregate the statistics by going through
all the items of each cpu sequentially which is reducing cache
misses.

Docker creation got faster by more than 2x after the patch.

Result:
                       Before           After
Docker creation time   6.836s           3.25s
cache miss             2.7%             1.41%

perf before:
    50.73%  docker           [kernel.kallsyms]       [k] snmp_fold_field
     9.07%  swapper          [kernel.kallsyms]       [k] snooze_loop
     3.49%  docker           [kernel.kallsyms]       [k] veth_stats_one
     2.85%  swapper          [kernel.kallsyms]       [k] _raw_spin_lock

perf after:
    10.57%  docker           docker                [.] scanblock
     8.37%  swapper          [kernel.kallsyms]     [k] snooze_loop
     6.91%  docker           [kernel.kallsyms]     [k] snmp_get_cpu_field
     6.67%  docker           [kernel.kallsyms]     [k] veth_stats_one

changes/ideas suggested:
Using buffer in stack (Eric), Usage of memset (David), Using memcpy in
place of unaligned_put (Joe).

Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv6/addrconf.c