diff options
author | Jon Marius Venstad <venstad@gmail.com> | 2021-08-18 10:45:00 +0200 |
---|---|---|
committer | Jon Marius Venstad <venstad@gmail.com> | 2021-08-18 10:45:00 +0200 |
commit | d1a979fc2aeee0df600bcb6bca5f643e8e71e9d5 (patch) | |
tree | b175c212f41f63ea7e68394a8dd3e2187e8ff05e /messagebus/src | |
parent | 5b61adcd248e9bd9f191c21c6d0a6dc39cf78d60 (diff) |
Cut the heaviest dynamic throttle policy mini-benchmarks
Diffstat (limited to 'messagebus/src')
-rw-r--r-- | messagebus/src/test/java/com/yahoo/messagebus/DynamicThrottlePolicyTest.java | 8 |
1 files changed, 4 insertions, 4 deletions
diff --git a/messagebus/src/test/java/com/yahoo/messagebus/DynamicThrottlePolicyTest.java b/messagebus/src/test/java/com/yahoo/messagebus/DynamicThrottlePolicyTest.java index 63747803e75..09fb1110fea 100644 --- a/messagebus/src/test/java/com/yahoo/messagebus/DynamicThrottlePolicyTest.java +++ b/messagebus/src/test/java/com/yahoo/messagebus/DynamicThrottlePolicyTest.java @@ -97,7 +97,7 @@ public class DynamicThrottlePolicyTest { /** Sort of a dummy test, as the conditions are perfect. In a more realistic scenario, below, the algorithm needs luck to climb this high. */ @Test public void singlePolicySingleWorkerWithIncreasingParallelism() { - for (int i = 0; i < 5; i++) { + for (int i = 0; i < 4; i++) { CustomTimer timer = new CustomTimer(); DynamicThrottlePolicy policy = new DynamicThrottlePolicy(timer); int scaleFactor = (int) Math.pow(10, i); @@ -120,11 +120,11 @@ public class DynamicThrottlePolicyTest { /** A more realistic test, where throughput gradually flattens with increasing window size, and with more variance in throughput. */ @Test public void singlePolicyIncreasingWorkersWithNoParallelism() { - for (int i = 0; i < 5; i++) { + for (int i = 0; i < 4; i++) { CustomTimer timer = new CustomTimer(); DynamicThrottlePolicy policy = new DynamicThrottlePolicy(timer); int scaleFactor = (int) Math.pow(10, i); - long operations = 5_000L * scaleFactor; + long operations = 2_000L * scaleFactor; // workPerSuccess determines the latency of the simulated server, which again determines the impact of the // synthetic attractors of the algorithm, around latencies which give (close to) integer log10(1 / latency). // With a value of 5, the impact is that the algorithm is pushed upwards slightly above 10k window size, @@ -143,7 +143,7 @@ public class DynamicThrottlePolicyTest { double maxMaxPending = numberOfWorkers * maximumTasksPerWorker; assertInRange(minMaxPending, summary.averagePending, maxMaxPending); assertInRange(minMaxPending, summary.averageWindows[0], maxMaxPending); - assertInRange(1, summary.inefficiency, 1 + 0.2 * i); // Even slower ramp-up. + assertInRange(1, summary.inefficiency, 1 + 0.25 * i); // Even slower ramp-up. assertInRange(0, summary.waste, 0); } } |