Update GCSFilter benchmarks

kcalvinalvin · kcalvinalvin · commit e734228d8585 · 2022-05-22T14:17:15.000+09:00
Element count used in the GCSFilter benchmarks are increased to 100,000
from 10,000. Testing the benchmarks with different element counts showed
that a filter with 100,000 elements resulted in the same ns/op. This
this a desirable thing to have as it allows us to reason about how long
a single filter element takes to process, letting us easily calculate
how long a filter with N elements (where N &gt; 100,000) would take to
process.

GCSFilterConstruct benchmark is now called without batch. This makes
intra-bench results more intuitive as all benchmarks are in ns/op
instead of a custom unit. There are no downsides to this change as
testing showed that there is no observable difference in error rates
in the benchmarks when calling without batch.
diff --git a/src/bench/gcs_filter.cpp b/src/bench/gcs_filter.cpp
@@ -8,7 +8,12 @@
 static const GCSFilter::ElementSet GenerateGCSTestElements()
 {
     GCSFilter::ElementSet elements;
-    for (int i = 0; i < 10000; ++i) {
+
+    // Testing the benchmarks with different number of elements show that a filter
+    // with at least 100,000 elements results in benchmarks that have the same
+    // ns/op. This makes it easy to reason about how long (in nanoseconds) a single
+    // filter element takes to process.
+    for (int i = 0; i < 100000; ++i) {
         GCSFilter::Element element(32);
         element[0] = static_cast<unsigned char>(i);
         element[1] = static_cast<unsigned char>(i >> 8);
@@ -35,7 +40,7 @@ static void GCSFilterConstruct(benchmark::Bench& bench)
     auto elements = GenerateGCSTestElements();
 
     uint64_t siphash_k0 = 0;
-    bench.batch(elements.size()).unit("elem").run([&] {
+    bench.run([&]{
         GCSFilter filter({siphash_k0, 0, BASIC_FILTER_P, BASIC_FILTER_M}, elements);
 
         siphash_k0++;
