Skip to content

DCPerf benchmark suite for hyperscale cloud applications

License

Notifications You must be signed in to change notification settings

facebookresearch/DCPerf

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

50 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DCPerf

DCPerf is a benchmark suite designed to represent real-world hyperscale cloud applications in the datacenter. It is meant for hardware vendors, system software developers and researchers to evaluate their new products and conduct performance projection & modeling in a way that can better represent real-world production workload developed by internet application companies and run in hyperscale cloud datacenters.

Overview

Hyper-scale and cloud datacenter deployments constitute the largest market share of server deployments in the world today. Workloads developed by large-scale internet companies running in their datacenters have very different characteristics than those in high performance computing (HPC) or traditional enterprise market segments. Therefore, server deployment considerations, trade-offs and objectives for data center use cases are also significantly different from other market segments and require a different set of benchmarks and evaluation methodology. Existing benchmarks (e.g. SPEC and Geekbench) fall short of capturing these characteristics and hence do not provide a reliable avenue to design and optimize modern server and datacenter designs.

To better evaluate the performance and efficiency of hardware for datacenter deployments, we developed DCPerf, a collection of benchmarks to represent the largest categories of workloads that run in cloud deployments. Each benchmark within DCPerf is designed by referencing a major application running in the cloud and then using numerous open-source frameworks and libraries with proper composition tuning to match the characteristics of that reference application. The motivation is that if we design and optimize the hardware and software on future server platforms for these benchmarks, it would help improve performance and efficiency of real-world cloud workloads.

For more information, see:

Workload Coverage

As of 2024 Q3, DCPerf consists of six benchmarks and provides coverage for the major production workloads listed as follows:

Benchmarks Programming Languages LIbraries / SW Stack Application domain they represent
Mediawiki PHP, Hack HHVM-3.30, Mediawiki, Memcached, MySQL, Nginx, Siege Web Serving (Facebook)
FeedSim C++ Oldisim Library, ZLIB, Boost, OpenSSL, BZIP2, LZ4, Snappy, libevent, jemalloc, lzma, libsodium, rsocket, fmt, FBThrift, Folly, wangle, fizz Object Aggregation, Ranking/Inference
TaoBench C++ Memcached, Memtier, Folly Caching, Look-through Cache
SparkBench Java, SQL Apache Spark, OpenJDK Data Analytics, Query Engine
DjangoBench Python, C++ Django framework, UWSGI, Apache Cassandra, Memcached, Siege Web Serving (Instagram)
VideoTranscodeBench C++ ffmpeg, svt-av1, libaom, x264 Video Processing

Also, DCPerf includes micro-benchmarks focusing on hot and commons functions/routines in the datacenter.

Note: WDLBench has different usages than other benchmarks. To use it, please first go through this README, and then go to WDLBench's specific README.

Benchmarks Programming Languages LIbraries / SW Stack Application domain they represent
WDLBench C++ folly, zstd, openssl Widely distributed functions across different workloads

Representativeness

When designing DCPerf, our goal is to have it represent real-world production workloads. In other words, micro-architecture and system optimizations that show improvement in DCPerf benchmark scores can potentially help improve performance of production workloads. We establish such representativeness in three levels:

  • Gen-over-gen Performance Improvment: The extent of performance improvement on a newer, more powerful compute platform seen in DCPerf benchmarks is close to the improvement that will show in the production application.
  • Hot Function Composition: DCPerf benchmarks have the similar types of top hot functions as their corresponding production applications, and the ratio among the hot functions will also roughly match.
  • Micro-architecture Metrics and Top-down Analysis: DCPerf can represent the most important micro-architecture and top-down metrics that the production applications show bottleneck and need optimization.

Versioning

DCPerf adopts semantics versioning for its releases:

  • Version number is represented as X.Y.Z, where X denotes MAJOR version, Y denotes MINOR version and Z denotes PATCH version
  • The initial release will have the version number of 0.1.0
  • Increment PATCH version when we make bug fixes or small enhancements to existing features (e.g. additional benchmark options, improvements to Benchpress framework or Perf hook)
  • Increment MINOR version when we make larger new changes but does not break backward compatibility, for example but not limited to:
    • Scalability improvements in benchmarks that improve scores & utilization on newer CPUs but do not change scores on older ones;
    • Introducing new features in Benchpress framework
    • Introducing new metrics / new CPU support in the Perf hook
    • Releasing new benchmarks to DCPerf
  • Increment MAJOR version when we make substantial changes to the suite which will make scores obtained from this version incomparable to the older versions.

Getting Started

This section will guide you through how to setup DCPerf, run benchmarks, collect scores, monitor performance and identify whether a benchmark run is good. For quick navigation, please feel free to use the following Table of Contents:

System Requirements

  • CPU Architecture: x86_64 or aarch64
  • OS: CentOS Stream 8/9, Ubuntu 22.04
  • Running as the root user
  • Have access to the internet
  • Please set ulimit -n to at least 65536. For permanent change please edit /etc/security/limits.conf

Install Prerequisites

On CentOS 8

Install git, python (>= 3.7) and the following Python packages:

  • click
  • pyyaml
  • tabulate
  • pandas

The commands are:

dnf install -y python38 python38-pip git
alternatives --set python3 /usr/bin/python3.8
pip-3.8 install click pyyaml tabulate pandas

Enable EPEL and PowerTools repos:

dnf install epel-release
dnf install 'dnf-command(config-manager)'
dnf config-manager --set-enabled PowerTools

After that, try running ./benchpress_cli.py under the benchpress directory.

NOTE: Since CentOS Stream 8 has reached EOL as of June 2024, some DCPerf's dependencies (such as folly) may start to drop its support. You may also encounter some troubles when trying to install packages via dnf. Therefore we recommend upgrading your OS to CentOS Stream 9. The newer version of folly may have begun to require newer versions of GCC compilers. If you still wish to run DCPerf on CentOS Stream 8, please install and enable GCC 11 with the following steps:

dnf install -y gcc-toolset-11
scl enable gcc-toolset-11 bash

On CentOS 9

Install git, click, pyyaml and tabulate using DNF, then install Pandas with pip:

dnf install -y git python3-click python3-pyyaml python3-tabulate python3-pip
pip-3.9 install pandas

Enable EPEL and PowerTools/CRB repos:

dnf install epel-release
dnf install 'dnf-command(config-manager)'
dnf config-manager --set-enabled CRB

On Ubuntu 22.04

Install git, pip, then install click, pyyaml, tabulate and pandas:

sudo apt update
sudo apt install -y python3-pip git
sudo pip3 install click pyyaml tabulate pandas

Using Benchpress

DCPerf uses Benchpress as the driver and framework to install, execute benchmarks, report results, collect system information and monitor performance telemetries.

After installing the aforementioned prerequisite packages, Benchpress CLI should work. You can test by trying the following command:

./benchpress_cli.py list

This command should list all currently available benchmark jobs in DCPerf, for example:

Job                              Tags                  Description
oss_performance_mediawiki        app,web,cpu           Default run for oss_performance_mediawiki
oss_performance_mediawiki_mlp    app,web,cpu           Tuned +MLP run for oss_performance_mediawiki
oss_performance_mediawiki_mem    app,web,cpu           Tuned +(MLP+LambdaChase) run for oss_performance_mediawiki
django_workload_default          app,django,cpu        Default run for django-workload
django_workload_arm              app,django,cpu        django-workload workload for ARM
django_workload_custom           app,django,cpu        Django-workload benchmark with custom parameters
......

You can also view the detailed information of a particular benchmark job using the info command followed by the job name. For example:

root@hostname:DCPerf/ $ ./benchpress_cli.py info django_workload_default
Properties    Values
--- Job ---   django_workload_default
Description   Default run for django-workload
Roles         'clientserver, db'
Arguments     {'clientserver': {'args': ['-r clientserver',
                                         '-d {duration}',
                                         '-i {iterations}',
                                         '-p {reps}',
                                         '-l ./siege.log',
                                         '-s urls.txt',
                                         '-c {db_addr}'],
                                'vars': ['db_addr', 'duration=5M', 'iterations=7', 'reps=0']},
               'db': {'args': ['-r db']}}
Hooks         [{'hook': 'copymove',
                'options': {'after': ['benchmarks/django_workload/django-workload/client/perf.data'],
                            'is_move': True}}]

This allows you to learn what roles does the benchmark job have and what parameters the benchmark job accepts.

Install and Run Benchmarks

NOTE: Each DCPerf benchmark has its own instruction for configuration and execution. Please click on the links below to view the detailed instructions on setting up and running the benchmarks. What's discussed in this section is an overview of Benchpress's install and run commands.

Installation

Before running a benchmark you need to install it. To install a benchmark in DCPerf, you can simply run ./benchpress_cli.py install <job_name> command. The installation process will download and install the required third-party dependencies and build the benchmark. For example:

./benchpress_cli.py install tao_bench_autoscale

If you have installed a benchmark before but would like to re-install, please add -f flag:

./benchpress_cli.py install -f tao_bench_autoscale

Note that you will need to re-install the benchmark to make updates to the benchmarks apply when you pull a newer version of DCPerf from Github. For cleaner re-install, it's recommended to uninstall first using the clean command

Uninstall

To uninstall a benchmark, you can run the clean command. For example:

./benchpress_cli.py clean tao_bench_autoscale

Basically what it does is to remove the artifacts under benchmarks/<bm_name>, which usually includes the executable of the benchmarks and some locally built third party libraries. It will not remove dependencies installed manually (such as HHVM) and installed to the system through dnf or apt.

Execution

To run a benchmark, you can use Benchpress's run command. For example:

./benchpress_cli.py run <benchmark_job_name> [-r role] [-i input_args]

There are two optional arguments, whether they are needed depend on the particular benchmark you want to run, so please refer to the detailed instructions linked above.

  • -r role: The role in the benchmark to run. For example, server, client.
  • -i input_args: The user input parameters for the benchmark job. input_args should be a quoted JSON string (e.g. -i '{"key1": "value1", "key2": "value2"}')

Getting results

After the benchmark is successfully run, it will print out a JSON object containing some basic information of the systems, benchmark specifications and benchmark results. You may check out one of the detailed benchmark instructions linked above to see the sample output.

Besides, Benchpress will create a folder called benchmark_metrics_<run_id> for the benchmark results. <run_id> is the UUID that will be printed out when the benchmark successfully finishes. Inside the folder there will be at least two files:

  • <benchmark_job_name>_metrics_<timestamp>_iter_None.json: This will contain the same JSON data that Benchpress outputs when benchmark finishes;
  • <benchmark_job_name>_system_specs_<timestamp>.json: This will record detailed system configurations of the machine, such as:
    • CPU topology
    • DMI
    • Hardware
    • Kernel
    • OS Distro
    • Installed packages

Getting DCPerf Score

After running five DCPerf benchmarks (TaoBench, FeedSim, DjangoBench, Mediawiki, SparkBench), you can obtain an overall DCPerf score of your test machine by running ./benchpress_cli.py report score command. We are currently in the process of validating correlation of the Video Transcode benchmark so it's not yet included in DCPerf overall score calculation. For example:

[root@hostname ~/DCPerf]# ./benchpress_cli.py report score
mediawiki: 4.741, single data point
django: 4.871, single data point
feedsim: 5.842, single data point
sparkbench: 3.361, single data point
taobench: 4.041, single data point
DCPerf overall score: 4.494

Note that by default the score reporter uses the latest results of the five benchmarks. You can add the --all flag to have the reporter use all benchmark results available in the DCPerf folder:

[root@hostname ~/DCPerf]# ./benchpress_cli.py report score --all
mediawiki: 5.211, median of 16 data points, stdev 4.83%, mean 5.067
django: 5.548, median of 4 data points, stdev 0.24%, mean 5.552
feedsim: 6.596, median of 4 data points, stdev 0.41%, mean 6.585
sparkbench: 3.620, median of 4 data points, stdev 0.51%, mean 3.622
taobench: 4.176, median of 4 data points, stdev 1.40%, mean 4.203
DCPerf overall score: 4.920

When there are multiple datapoints for a benchmark, the reporter will report median as the score, followed by how many data points detected, run-to-run variation and the mean/average. If there are only two datapoints, the reporter will report the average of the two.

When not all the five DCPerf benchmarks have been run, the reporter will also give a geometric mean score based off what's available there, but it will be marked as "partial", meaning it could not be treated as a complete overall score:

[root@hostname ~/DCPerf]# ./benchpress_cli.py report score --all
feedsim: 6.596, median of 4 data points, stdev 0.41%, mean 6.585
sparkbench: 3.620, median of 4 data points, stdev 0.51%, mean 3.622
taobench: 4.176, median of 4 data points, stdev 1.40%, mean 4.203
DCPerf partial geomean: 4.637

The score is defined as follows:

  1. Each benchmark has a baseline performance metric, which is basically the result we obtained from running DCPerf on a reference machine.
  2. The "score" of a benchmark is defined as the ratio of the performance metric achieved by the test machine to the baseline.
  3. The DCPerf overall score is the geometric mean of all five benchmarks' scores.

Monitoring system performance metrics

Benchpress provides a hook called perf that can help you monitor system performance metrics such as CPU utilization, memory usage, CPU frequency, network bandwidth and some micro-architecture telemetries while running DCPerf benchmarks.

Regarding how to use this hook and what functionalities it can provide, please refer to this README.

Expected CPU Utilization

Below is a table of our expectation on CPU utilization of these benchmarks. They can be used as a reference to see if a benchmark has run successfully and sufficiently stressed the system.

Benchmark Criteria CPU Utilization Target
TaoBench Last 5~10 minutes (determined by `test_time` parameter) 70~80% overall, 15~20% user
FeedSim Last 5 minutes 60~75%
DjangoWorkload Entire benchmark ~95%
Mediawiki Last 10 minutes 90~100%
SparkBench Entire benchmark 55~75%
SparkBench Stage 2.0 full batch period 90~100%
VideoTranscodeBench Encoding periods (most of the execution time) 85~100%

Limitations and Future Works

  1. Memory Representativeness. DCPerf benchmarks generally exert less memory bandwidth and capacity pressure on the test machines than the actual production workloads on servers. We will continue to work on evaluating and improving memory representativeness in these benchmarks.

  2. Software Optimization Reflection. When benchmarks in DCPerf were designed and implemented, they were referenced and evaluated based on earler versions of production applications. Therefore, they may not reflect software optimizations that occurred later on for some more modern ISAs. We will actively update these benchmarks to catch up with the constantly evolving production applications.

LICENSE

This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree.