While creating 1000 containers, perf is showing lot of time spent in snmp_fold_field on a large cpu system.
The current patch tries to improve by reordering the statistics gathering. Please note that similar overhead was also reported while creating veth pairs https://lkml.org/lkml/2013/3/19/556 Changes in V4: - remove 'item' variable and use IPSTATS_MIB_MAX to avoid sparse warning (Eric) also remove 'item' parameter (Joe) - add missing memset of padding. Changes in V3: - use memset to initialize temp buffer in leaf function. (David) - use memcpy to copy the buffer data to stat instead of unalign_pu (Joe) - Move buffer definition to leaf function __snmp6_fill_stats64() (Eric) - Changes in V2: - Allocate the stat calculation buffer in stack. (Eric) Setup: 160 cpu (20 core) baremetal powerpc system with 1TB memory 1000 docker containers was created with command docker run -itd ubuntu:15.04 /bin/bash in loop observation: Docker container creation linearly increased from around 1.6 sec to 7.5 sec (at 1000 containers) perf data showed, creating veth interfaces resulting in the below code path was taking more time. rtnl_fill_ifinfo -> inet6_fill_link_af -> inet6_fill_ifla6_attrs -> snmp_fold_field proposed idea: currently __snmp6_fill_stats64 calls snmp_fold_field that walks through per cpu data to of an item (iteratively for around 36 items). The patch tries to aggregate the statistics by going through all the items of each cpu sequentially which is reducing cache misses. Performance of docker creation improved by around more than 2x after the patch. before the patch: ================ #time docker run -itd ubuntu:15.04 /bin/bash 3f45ba571a42e925c4ec4aaee0e48d7610a9ed82a4c931f83324d41822cf6617 real 0m6.836s user 0m0.095s sys 0m0.011s perf record -a docker run -itd ubuntu:15.04 /bin/bash ======================================================= # Samples: 32K of event 'cycles' # Event count (approx.): 24688700190 # Overhead Command Shared Object Symbol # ........ ............... ...................... ........................ 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 1.37% docker docker [.] backtrace_qsort 1.31% docker docker [.] strings.FieldsFunc cache-misses: 2.7% after the patch: ============= #time docker run -itd ubuntu:15.04 /bin/bash 9178273e9df399c8290b6c196e4aef9273be2876225f63b14a60cf97eacfafb5 real 0m3.249s user 0m0.088s sys 0m0.020s perf record -a docker run -itd ubuntu:15.04 /bin/bash ======================================================= # Samples: 18K of event 'cycles' # Event count (approx.): 13958958797 # Overhead Command Shared Object Symbol # ........ ............... .................... .......................... 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 3.96% docker docker [.] runtime_MSpan_Sweep 2.47% docker docker [.] strings.FieldsFunc cache-misses: 1.41 % Please let me know if you have suggestions/comments. Thanks Eric, Joe and David for the comments. Raghavendra K T (2): net: Introduce helper functions to get the per cpu data net: Optimize snmp stat aggregation by walking all the percpu data at once include/net/ip.h | 10 ++++++++++ net/ipv4/af_inet.c | 41 +++++++++++++++++++++++++++-------------- net/ipv6/addrconf.c | 26 ++++++++++++++++---------- 3 files changed, 53 insertions(+), 24 deletions(-) -- 1.7.11.7 -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
