pkg/registry/pod: avoid allocation in common pod search

PodToSelectableFields creates a map of field attributes
for a particular pod filter query to use. If the result
of the query does not depend on the fields at all, avoid
creating the map.

This is the source of about half the allocated memory
(by byte volume) during the kubemark benchmark, and it
is in turn the main driver of CPU usage during the benchmark,
because of the many background pod watches going on,
as well as the occasional list pods.

These benchmarks for 1000-node kubemark show the difference
from my previous CL (caching timers) to this CL:

name                               old ms/op   new ms/op   delta
LIST_nodes_p50                       124 ±13%    121 ± 9%     ~     (p=0.136 n=29+27)
LIST_nodes_p90                       278 ±15%    266 ±12%   -4.26%  (p=0.031 n=29+27)
LIST_nodes_p99                       405 ±19%    400 ±14%     ~     (p=0.864 n=28+28)
LIST_pods_p50                       65.3 ±13%   56.3 ± 9%  -13.75%  (p=0.000 n=29+28)
LIST_pods_p90                        115 ±12%     93 ± 8%  -18.75%  (p=0.000 n=27+28)
LIST_pods_p99                        226 ±21%    202 ±14%  -10.52%  (p=0.000 n=28+28)
LIST_replicationcontrollers_p50     26.6 ±43%   26.2 ±54%     ~     (p=0.487 n=29+29)
LIST_replicationcontrollers_p90     68.7 ±63%   68.6 ±59%     ~     (p=0.931 n=29+28)
LIST_replicationcontrollers_p99      173 ±41%    177 ±49%     ~     (p=0.618 n=28+29)
PUT_replicationcontrollers_p50      5.83 ±36%   5.94 ±32%     ~     (p=0.818 n=28+29)
PUT_replicationcontrollers_p90      15.9 ± 6%   15.5 ± 6%   -2.23%  (p=0.019 n=28+29)
PUT_replicationcontrollers_p99      56.7 ±41%   39.5 ±55%  -30.29%  (p=0.000 n=28+29)
DELETE_pods_p50                     24.3 ±17%   24.3 ±13%     ~     (p=0.855 n=28+29)
DELETE_pods_p90                     30.6 ± 0%   30.7 ± 1%     ~     (p=0.140 n=28+29)
DELETE_pods_p99                     56.3 ±27%   54.2 ±23%     ~     (p=0.188 n=28+27)
PUT_nodes_p50                       14.9 ± 1%   14.8 ± 2%     ~     (p=0.781 n=28+27)
PUT_nodes_p90                       16.4 ± 2%   16.3 ± 2%     ~     (p=0.321 n=28+28)
PUT_nodes_p99                       44.6 ±42%   41.3 ±35%     ~     (p=0.361 n=29+28)
POST_replicationcontrollers_p50     6.33 ±23%   6.34 ±20%     ~     (p=0.993 n=28+28)
POST_replicationcontrollers_p90     15.2 ± 6%   15.0 ± 5%     ~     (p=0.106 n=28+29)
POST_replicationcontrollers_p99     53.4 ±52%   32.9 ±46%  -38.41%  (p=0.000 n=27+27)
POST_pods_p50                       9.33 ±13%   8.95 ±16%     ~     (p=0.069 n=29+29)
POST_pods_p90                       16.3 ± 4%   16.1 ± 4%   -1.43%  (p=0.044 n=29+29)
POST_pods_p99                       28.4 ±23%   26.4 ±12%   -7.05%  (p=0.004 n=29+28)
DELETE_replicationcontrollers_p50   2.50 ±13%   2.50 ±13%     ~     (p=0.649 n=29+28)
DELETE_replicationcontrollers_p90   11.7 ±10%   11.8 ±13%     ~     (p=0.863 n=28+28)
DELETE_replicationcontrollers_p99   19.0 ±22%   19.1 ±21%     ~     (p=0.818 n=28+29)
PUT_pods_p50                        10.3 ± 5%   10.2 ± 5%     ~     (p=0.235 n=28+27)
PUT_pods_p90                        16.0 ± 1%   16.0 ± 1%     ~     (p=0.380 n=29+28)
PUT_pods_p99                        21.6 ±14%   20.9 ± 9%   -3.15%  (p=0.010 n=28+27)
POST_bindings_p50                   8.98 ±17%   8.92 ±15%     ~     (p=0.666 n=29+28)
POST_bindings_p90                   16.5 ± 2%   16.5 ± 3%     ~     (p=0.840 n=26+29)
POST_bindings_p99                   21.4 ± 5%   21.1 ± 4%   -1.21%  (p=0.049 n=27+28)
GET_nodes_p90                       1.18 ±19%   1.14 ±24%     ~     (p=0.137 n=29+29)
GET_nodes_p99                       8.29 ±40%   7.50 ±46%     ~     (p=0.106 n=28+29)
GET_replicationcontrollers_p90      1.03 ±21%   1.01 ±27%     ~     (p=0.489 n=29+29)
GET_replicationcontrollers_p99     10.0 ±123%  10.0 ±145%     ~     (p=0.794 n=28+29)
GET_pods_p90                        1.08 ±21%   1.02 ±19%     ~     (p=0.083 n=29+28)
GET_pods_p99                        2.81 ±39%   2.45 ±38%  -12.78%  (p=0.021 n=28+25)

Overall the two CLs combined have this effect:

name                               old ms/op  new ms/op   delta
LIST_nodes_p50                      127 ±16%    121 ± 9%   -4.58%  (p=0.000 n=29+27)
LIST_nodes_p90                      326 ±12%    266 ±12%  -18.48%  (p=0.000 n=29+27)
LIST_nodes_p99                      453 ±11%    400 ±14%  -11.79%  (p=0.000 n=29+28)
LIST_replicationcontrollers_p50    29.4 ±49%   26.2 ±54%     ~     (p=0.085 n=30+29)
LIST_replicationcontrollers_p90    83.0 ±78%   68.6 ±59%  -17.33%  (p=0.013 n=30+28)
LIST_replicationcontrollers_p99     216 ±43%    177 ±49%  -17.68%  (p=0.000 n=29+29)
DELETE_pods_p50                    24.5 ±14%   24.3 ±13%     ~     (p=0.562 n=30+29)
DELETE_pods_p90                    30.7 ± 1%   30.7 ± 1%   -0.30%  (p=0.011 n=29+29)
DELETE_pods_p99                    77.2 ±34%   54.2 ±23%  -29.76%  (p=0.000 n=30+27)
PUT_replicationcontrollers_p50     5.86 ±26%   5.94 ±32%     ~     (p=0.734 n=29+29)
PUT_replicationcontrollers_p90     15.8 ± 7%   15.5 ± 6%   -2.06%  (p=0.010 n=29+29)
PUT_replicationcontrollers_p99     57.8 ±35%   39.5 ±55%  -31.60%  (p=0.000 n=29+29)
PUT_nodes_p50                      14.9 ± 2%   14.8 ± 2%   -0.68%  (p=0.012 n=30+27)
PUT_nodes_p90                      16.5 ± 1%   16.3 ± 2%   -0.90%  (p=0.000 n=27+28)
PUT_nodes_p99                      57.9 ±47%   41.3 ±35%  -28.61%  (p=0.000 n=30+28)
POST_replicationcontrollers_p50    6.35 ±29%   6.34 ±20%     ~     (p=0.944 n=30+28)
POST_replicationcontrollers_p90    15.4 ± 5%   15.0 ± 5%   -2.18%  (p=0.001 n=29+29)
POST_replicationcontrollers_p99    52.2 ±71%   32.9 ±46%  -36.99%  (p=0.000 n=29+27)
POST_pods_p50                      8.99 ±13%   8.95 ±16%     ~     (p=0.903 n=30+29)
POST_pods_p90                      16.2 ± 4%   16.1 ± 4%     ~     (p=0.287 n=29+29)
POST_pods_p99                      30.9 ±21%   26.4 ±12%  -14.73%  (p=0.000 n=28+28)
POST_bindings_p50                  9.34 ±12%   8.92 ±15%   -4.54%  (p=0.013 n=30+28)
POST_bindings_p90                  16.6 ± 1%   16.5 ± 3%   -0.73%  (p=0.017 n=28+29)
POST_bindings_p99                  23.5 ± 9%   21.1 ± 4%  -10.09%  (p=0.000 n=27+28)
PUT_pods_p50                       10.8 ±11%   10.2 ± 5%   -5.47%  (p=0.000 n=30+27)
PUT_pods_p90                       16.1 ± 1%   16.0 ± 1%   -0.64%  (p=0.000 n=29+28)
PUT_pods_p99                       23.4 ± 9%   20.9 ± 9%  -10.93%  (p=0.000 n=28+27)
DELETE_replicationcontrollers_p50  2.42 ±16%   2.50 ±13%     ~     (p=0.054 n=29+28)
DELETE_replicationcontrollers_p90  11.5 ±12%   11.8 ±13%     ~     (p=0.141 n=30+28)
DELETE_replicationcontrollers_p99  19.5 ±21%   19.1 ±21%     ~     (p=0.397 n=29+29)
GET_nodes_p50                      0.77 ±10%   0.76 ±10%     ~     (p=0.317 n=28+28)
GET_nodes_p90                      1.20 ±16%   1.14 ±24%   -4.66%  (p=0.036 n=28+29)
GET_nodes_p99                      11.4 ±48%    7.5 ±46%  -34.28%  (p=0.000 n=28+29)
GET_replicationcontrollers_p50     0.74 ±17%   0.73 ±17%     ~     (p=0.222 n=30+28)
GET_replicationcontrollers_p90     1.04 ±25%   1.01 ±27%     ~     (p=0.231 n=30+29)
GET_replicationcontrollers_p99     12.1 ±81%  10.0 ±145%     ~     (p=0.063 n=28+29)
GET_pods_p50                       0.78 ±12%   0.77 ±10%     ~     (p=0.178 n=30+28)
GET_pods_p90                       1.06 ±19%   1.02 ±19%     ~     (p=0.120 n=29+28)
GET_pods_p99                       3.92 ±43%   2.45 ±38%  -37.55%  (p=0.000 n=27+25)
LIST_services_p50                  0.20 ±13%   0.20 ±16%     ~     (p=0.854 n=28+29)
LIST_services_p90                  0.28 ±15%   0.27 ±14%     ~     (p=0.219 n=29+28)
LIST_services_p99                  0.49 ±20%   0.47 ±24%     ~     (p=0.140 n=29+29)
LIST_endpoints_p50                 0.19 ±14%   0.19 ±15%     ~     (p=0.709 n=29+29)
LIST_endpoints_p90                 0.26 ±16%   0.26 ±13%     ~     (p=0.274 n=29+28)
LIST_endpoints_p99                 0.46 ±24%   0.44 ±21%     ~     (p=0.111 n=29+29)
LIST_horizontalpodautoscalers_p50  0.16 ±15%   0.15 ±13%     ~     (p=0.253 n=30+27)
LIST_horizontalpodautoscalers_p90  0.22 ±24%   0.21 ±16%     ~     (p=0.152 n=30+28)
LIST_horizontalpodautoscalers_p99  0.31 ±33%   0.31 ±38%     ~     (p=0.817 n=28+29)
LIST_daemonsets_p50                0.16 ±20%   0.15 ±11%     ~     (p=0.135 n=30+27)
LIST_daemonsets_p90                0.22 ±18%   0.21 ±25%     ~     (p=0.135 n=29+28)
LIST_daemonsets_p99                0.29 ±28%   0.29 ±32%     ~     (p=0.606 n=28+28)
LIST_jobs_p50                      0.16 ±16%   0.15 ±12%     ~     (p=0.375 n=29+28)
LIST_jobs_p90                      0.22 ±18%   0.21 ±16%     ~     (p=0.090 n=29+26)
LIST_jobs_p99                      0.31 ±28%   0.28 ±35%  -10.29%  (p=0.005 n=29+27)
LIST_deployments_p50               0.15 ±16%   0.15 ±13%     ~     (p=0.565 n=29+28)
LIST_deployments_p90               0.22 ±22%   0.21 ±19%     ~     (p=0.107 n=30+28)
LIST_deployments_p99               0.31 ±27%   0.29 ±34%     ~     (p=0.068 n=29+28)
LIST_namespaces_p50                0.21 ±25%   0.21 ±26%     ~     (p=0.768 n=29+27)
LIST_namespaces_p90                0.28 ±29%   0.26 ±25%     ~     (p=0.101 n=30+28)
LIST_namespaces_p99                0.30 ±48%   0.29 ±42%     ~     (p=0.339 n=30+29)
LIST_replicasets_p50               0.15 ±18%   0.15 ±16%     ~     (p=0.612 n=30+28)
LIST_replicasets_p90               0.22 ±19%   0.21 ±18%   -5.13%  (p=0.011 n=28+27)
LIST_replicasets_p99               0.31 ±39%   0.28 ±29%     ~     (p=0.066 n=29+28)
LIST_persistentvolumes_p50         0.16 ±23%   0.15 ±21%     ~     (p=0.124 n=30+29)
LIST_persistentvolumes_p90         0.21 ±23%   0.20 ±23%     ~     (p=0.092 n=30+25)
LIST_persistentvolumes_p99         0.21 ±24%   0.20 ±23%     ~     (p=0.053 n=30+25)
LIST_resourcequotas_p50            0.16 ±12%   0.16 ±13%     ~     (p=0.175 n=27+28)
LIST_resourcequotas_p90            0.20 ±22%   0.20 ±24%     ~     (p=0.388 n=30+28)
LIST_resourcequotas_p99            0.22 ±24%   0.22 ±23%     ~     (p=0.575 n=30+28)
LIST_persistentvolumeclaims_p50    0.15 ±21%   0.15 ±29%     ~     (p=0.079 n=30+28)
LIST_persistentvolumeclaims_p90    0.19 ±26%   0.18 ±34%     ~     (p=0.446 n=29+29)
LIST_persistentvolumeclaims_p99    0.19 ±26%   0.18 ±34%     ~     (p=0.446 n=29+29)
LIST_pods_p50                      68.0 ±16%   56.3 ± 9%  -17.19%  (p=0.000 n=29+28)
LIST_pods_p90                       119 ±19%     93 ± 8%  -21.88%  (p=0.000 n=28+28)
LIST_pods_p99                       230 ±18%    202 ±14%  -12.13%  (p=0.000 n=27+28)
This commit is contained in:
Russ Cox 2016-04-04 13:29:52 -04:00
parent 6a19e46ed6
commit 58629a28e4

View File

@ -163,7 +163,19 @@ func MatchPod(label labels.Selector, field fields.Selector) generic.Matcher {
if !ok {
return nil, nil, fmt.Errorf("not a pod")
}
return labels.Set(pod.ObjectMeta.Labels), PodToSelectableFields(pod), nil
// podLabels is already sitting there ready to be used.
// podFields is not available directly and requires allocation of a map.
// Only bother if the fields might be useful to determining the match.
// One common case is for a replication controller to set up a watch
// based on labels alone; in that case we can avoid allocating the field map.
// This is especially important in the apiserver.
podLabels := labels.Set(pod.ObjectMeta.Labels)
var podFields fields.Set
if !field.Empty() && label.Matches(podLabels) {
podFields = PodToSelectableFields(pod)
}
return podLabels, podFields, nil
},
}
}