2020-07-28 19:51:54 +00:00
|
|
|
// Copyright 2020 The gVisor Authors.
|
|
|
|
//
|
|
|
|
// Licensed under the Apache License, Version 2.0 (the "License");
|
|
|
|
// you may not use this file except in compliance with the License.
|
|
|
|
// You may obtain a copy of the License at
|
|
|
|
//
|
|
|
|
// http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
//
|
|
|
|
// Unless required by applicable law or agreed to in writing, software
|
|
|
|
// distributed under the License is distributed on an "AS IS" BASIS,
|
|
|
|
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
|
|
// See the License for the specific language governing permissions and
|
|
|
|
// limitations under the License.
|
|
|
|
package ml
|
|
|
|
|
|
|
|
import (
|
|
|
|
"context"
|
2020-12-09 19:55:06 +00:00
|
|
|
"os"
|
2020-07-28 19:51:54 +00:00
|
|
|
"testing"
|
|
|
|
|
|
|
|
"gvisor.dev/gvisor/pkg/test/dockerutil"
|
|
|
|
"gvisor.dev/gvisor/test/benchmarks/harness"
|
|
|
|
)
|
|
|
|
|
|
|
|
// BenchmarkTensorflow runs workloads from a TensorFlow tutorial.
|
|
|
|
// See: https://github.com/aymericdamien/TensorFlow-Examples
|
|
|
|
func BenchmarkTensorflow(b *testing.B) {
|
|
|
|
workloads := map[string]string{
|
|
|
|
"GradientDecisionTree": "2_BasicModels/gradient_boosted_decision_tree.py",
|
|
|
|
"Kmeans": "2_BasicModels/kmeans.py",
|
|
|
|
"LogisticRegression": "2_BasicModels/logistic_regression.py",
|
|
|
|
"NearestNeighbor": "2_BasicModels/nearest_neighbor.py",
|
|
|
|
"RandomForest": "2_BasicModels/random_forest.py",
|
|
|
|
"ConvolutionalNetwork": "3_NeuralNetworks/convolutional_network.py",
|
|
|
|
"MultilayerPerceptron": "3_NeuralNetworks/multilayer_perceptron.py",
|
|
|
|
"NeuralNetwork": "3_NeuralNetworks/neural_network.py",
|
|
|
|
}
|
|
|
|
|
Simplify profiling and benchmarks.
- Tweak the benchmarks to work with b.N where appropriate. In many cases,
b.N was simply being ignored. This creates an implicit dependency in the
user passing a reasonable benchtime (less than or equal to the actual
runtime of the test, or using the X syntax) otherwise the test runs
forever.
- In cases where the above is impossible, explicitly set benchtime from
the test wrapper, to prevent the above behavior (tensorflow).
- Drop the *Reverse variants, which are simply hey benchmarks. We should
just add a hey benchmark. The platforms benchmarks already include a
native platform, and thus these benchmarks are incredibly confusing.
(In other words, BenchmarkNginxReverse has nothing to do with an nginx
benchmark for runsc.)
- Remove the redunant Harness object, which contains no state, in order
to slightly simplify the code.
- Make Block and Heap profiling actually work, but setting appropriate
runtime parameters (and plumbing them through the config).
- Split the profiling into two phases: start and stop, since some will
need to be started early, and others will need to happen at the end.
PiperOrigin-RevId: 349495377
2020-12-30 02:26:46 +00:00
|
|
|
machine, err := harness.GetMachine()
|
2020-07-28 19:51:54 +00:00
|
|
|
if err != nil {
|
|
|
|
b.Fatalf("failed to get machine: %v", err)
|
|
|
|
}
|
|
|
|
defer machine.CleanUp()
|
|
|
|
|
|
|
|
for name, workload := range workloads {
|
|
|
|
b.Run(name, func(b *testing.B) {
|
|
|
|
ctx := context.Background()
|
|
|
|
|
|
|
|
b.ResetTimer()
|
Simplify profiling and benchmarks.
- Tweak the benchmarks to work with b.N where appropriate. In many cases,
b.N was simply being ignored. This creates an implicit dependency in the
user passing a reasonable benchtime (less than or equal to the actual
runtime of the test, or using the X syntax) otherwise the test runs
forever.
- In cases where the above is impossible, explicitly set benchtime from
the test wrapper, to prevent the above behavior (tensorflow).
- Drop the *Reverse variants, which are simply hey benchmarks. We should
just add a hey benchmark. The platforms benchmarks already include a
native platform, and thus these benchmarks are incredibly confusing.
(In other words, BenchmarkNginxReverse has nothing to do with an nginx
benchmark for runsc.)
- Remove the redunant Harness object, which contains no state, in order
to slightly simplify the code.
- Make Block and Heap profiling actually work, but setting appropriate
runtime parameters (and plumbing them through the config).
- Split the profiling into two phases: start and stop, since some will
need to be started early, and others will need to happen at the end.
PiperOrigin-RevId: 349495377
2020-12-30 02:26:46 +00:00
|
|
|
b.StopTimer()
|
|
|
|
|
2020-07-28 19:51:54 +00:00
|
|
|
for i := 0; i < b.N; i++ {
|
2020-12-09 19:55:06 +00:00
|
|
|
container := machine.GetContainer(ctx, b)
|
|
|
|
defer container.CleanUp(ctx)
|
2020-07-28 19:51:54 +00:00
|
|
|
if err := harness.DropCaches(machine); err != nil {
|
|
|
|
b.Skipf("failed to drop caches: %v. You probably need root.", err)
|
|
|
|
}
|
|
|
|
|
Simplify profiling and benchmarks.
- Tweak the benchmarks to work with b.N where appropriate. In many cases,
b.N was simply being ignored. This creates an implicit dependency in the
user passing a reasonable benchtime (less than or equal to the actual
runtime of the test, or using the X syntax) otherwise the test runs
forever.
- In cases where the above is impossible, explicitly set benchtime from
the test wrapper, to prevent the above behavior (tensorflow).
- Drop the *Reverse variants, which are simply hey benchmarks. We should
just add a hey benchmark. The platforms benchmarks already include a
native platform, and thus these benchmarks are incredibly confusing.
(In other words, BenchmarkNginxReverse has nothing to do with an nginx
benchmark for runsc.)
- Remove the redunant Harness object, which contains no state, in order
to slightly simplify the code.
- Make Block and Heap profiling actually work, but setting appropriate
runtime parameters (and plumbing them through the config).
- Split the profiling into two phases: start and stop, since some will
need to be started early, and others will need to happen at the end.
PiperOrigin-RevId: 349495377
2020-12-30 02:26:46 +00:00
|
|
|
// Run tensorflow.
|
|
|
|
b.StartTimer()
|
2020-07-28 19:51:54 +00:00
|
|
|
if out, err := container.Run(ctx, dockerutil.RunOpts{
|
|
|
|
Image: "benchmarks/tensorflow",
|
|
|
|
Env: []string{"PYTHONPATH=$PYTHONPATH:/TensorFlow-Examples/examples"},
|
|
|
|
WorkDir: "/TensorFlow-Examples/examples",
|
|
|
|
}, "python", workload); err != nil {
|
|
|
|
b.Fatalf("failed to run container: %v logs: %s", err, out)
|
|
|
|
}
|
Simplify profiling and benchmarks.
- Tweak the benchmarks to work with b.N where appropriate. In many cases,
b.N was simply being ignored. This creates an implicit dependency in the
user passing a reasonable benchtime (less than or equal to the actual
runtime of the test, or using the X syntax) otherwise the test runs
forever.
- In cases where the above is impossible, explicitly set benchtime from
the test wrapper, to prevent the above behavior (tensorflow).
- Drop the *Reverse variants, which are simply hey benchmarks. We should
just add a hey benchmark. The platforms benchmarks already include a
native platform, and thus these benchmarks are incredibly confusing.
(In other words, BenchmarkNginxReverse has nothing to do with an nginx
benchmark for runsc.)
- Remove the redunant Harness object, which contains no state, in order
to slightly simplify the code.
- Make Block and Heap profiling actually work, but setting appropriate
runtime parameters (and plumbing them through the config).
- Split the profiling into two phases: start and stop, since some will
need to be started early, and others will need to happen at the end.
PiperOrigin-RevId: 349495377
2020-12-30 02:26:46 +00:00
|
|
|
b.StopTimer()
|
2020-07-28 19:51:54 +00:00
|
|
|
}
|
|
|
|
})
|
|
|
|
}
|
|
|
|
}
|
2020-12-09 19:55:06 +00:00
|
|
|
|
|
|
|
func TestMain(m *testing.M) {
|
Simplify profiling and benchmarks.
- Tweak the benchmarks to work with b.N where appropriate. In many cases,
b.N was simply being ignored. This creates an implicit dependency in the
user passing a reasonable benchtime (less than or equal to the actual
runtime of the test, or using the X syntax) otherwise the test runs
forever.
- In cases where the above is impossible, explicitly set benchtime from
the test wrapper, to prevent the above behavior (tensorflow).
- Drop the *Reverse variants, which are simply hey benchmarks. We should
just add a hey benchmark. The platforms benchmarks already include a
native platform, and thus these benchmarks are incredibly confusing.
(In other words, BenchmarkNginxReverse has nothing to do with an nginx
benchmark for runsc.)
- Remove the redunant Harness object, which contains no state, in order
to slightly simplify the code.
- Make Block and Heap profiling actually work, but setting appropriate
runtime parameters (and plumbing them through the config).
- Split the profiling into two phases: start and stop, since some will
need to be started early, and others will need to happen at the end.
PiperOrigin-RevId: 349495377
2020-12-30 02:26:46 +00:00
|
|
|
harness.Init()
|
|
|
|
harness.SetFixedBenchmarks()
|
2020-12-09 19:55:06 +00:00
|
|
|
os.Exit(m.Run())
|
|
|
|
}
|