Currently Jade is not ready for open-source. However, we have set up a machine and installed a pre-compiled Jade binary executable on it. Additionally, we have prepared the corresponding testing application and automation scripts for reproducing the experiments described in the paper. The following is the directory structure of our artifact.
./jade-ae/
├── README.md
├── dacapo
├── hbase-2.4.14
├── specjbb2015
├── ycsb-hbase2-binding-0.18.0-SNAPSHOT
├── jdk
│ ├── jdk-11
│ ├── jdk-21-genz
│ ├── jdk-genshen
│ ├── jdk-jade
│ └── jdk-lxr
├── results
│ ├── dacapo
│ ├── specjbb
│ └── hbase
└── scripts
├── run-dacapo.sh
├── run-hbase.sh
└── run-specjbb.sh
Test applications including DaCapo, SPECjbb2015, HBase & YCSB can be found at ./jade-ae/root directory.
JDKs of Jade and other baselines can be found at ./jade-ae/jdk. This artifact includes jdk-jade(pre-compiled jdk with Jade), jdk11 (which is the base of Jade development), jdk-lxr (which is the state-of-the-art STW GC), jdk-genshen (a special version jdk supporting generational Shenandoah), and jdk-21-genz (the latest version of jdk21 supporting generational ZGC).
Scripts can be found at ./jade-ae/scripts and can be used to automatically run experiments, process outputs, and reproduce the results of our paper.
Results will be generated at ./jade-ae/results, and the following documentation will explain how to utilize these results.
Hardware dependencies. We provide an ECS instance on Alibaba Cloud to conduct all the experiments. The processor model of the ECS instance is Intel Xeon (Ice Lake) Platinum 8369B. Other x86_64 hardware can also use Jade and run our experiment.
Software dependencies. The operating system used in our machines is Alibaba Cloud Linux release 3, which is a distribution based on the 5.10.112-11.1.al8.x86_64 version of Linux kernel. But other distributions of Linux with higher kernel version are also acceptable.
Jade is a garbage collector developed based on JDK 11. You can use Jade just like using any other JDK.
$ ~/jade-ae/jdk/jdk-jade/bin/java --version
openjdk 11.0.17.13 2023-02-08
OpenJDK Runtime Environment (Alibaba Dragonwell) (build 11.0.17.13+0)
OpenJDK 64-Bit Server VM (Alibaba Dragonwell) (build 11.0.17.13+0, mixed mode)
You just need to add the parameters -XX:+UseJadeGC -XX:+JadeEnableChasingMode in the command to run Java applications with Jade. Below is an example of running the dacapo's lusearch with Jade.
$ ~/jade-ae/jdk/jdk-jade/bin/java -XX:+UseJadeGC -XX:+JadeEnableChasingMode -jar ~/jade-ae/dacapo/dacapo-evaluation-git-b00bfa9.jar lusearch
This artifact contains two major claims (C1, C2), which are proven by three experiments (E1-E3).
-
C1. In applications with low memory consumption, Jade exhibits slightly lower throughput compared to g1 and lxr, but outperforms other concurrent collectors significantly.
-
C2. In applications with higher memory usage, Jade demonstrates lower latency than other garbage collectors under high throughput conditions. While G1 is the only garbage collector capable of achieving slightly higher maximum throughput than Jade, this comes at the expense of noticeably higher latency.
C1
- 8 cores, about 4 hours to run all applications
C2
-
Specjbb2015: 8 cores, about 60 hours to run all data points
-
HBase: 32 cores, about 24 hours to run all points
| Experiment | Script | Table & Figure | Claim |
|---|---|---|---|
| E1 | ./run-dacapo.sh | Table 4 | C1 |
| E2 | ./run-specjbb.sh | Table 3 & Figure 4 | C2 |
| E3 | ./run-hbase.sh | Figure 5 | C2 |
Run the following script to evaluate baseline and jade.
$ cd ~/jade-ae/scripts
$ bash ./run-dacapo.sh
Results will be automatically generated at ~/jade-ae/results/dacapo/report.csv and should be consistent with Table 4.
Run the following script to evaluate baseline and jade.
$ cd ~/jade-ae/scripts
$ bash ./run-specjbb.sh
Reports will be automatically generated in the following folder: ~/jade-ae/results/specjbb
For example, the SPECjbb 2015 report with g1 and 4x heap size can be found at ~/jade-ae/results/specjbb/g1-heap4.0/report-00001/specjbb2015-C-XXXXXX-XXXXXX.html. The majority of the crucial data from the report is located at ~/jade-ae/results/specjbb/g1-heap4.0/report-00001/data/rt-curve/specjbb2015-C-XXXXXX-XXXXXX-overall-throughput-rt.txt. The maximum-jops and critical-jops of SPECjbb-g1-4xHeap testcase in Table 3 can be found at the beginning of this txt file, along with the throughput-p99 data of Figure 4 following it.
Run the following script to generate report.
$ cd ~/jade-ae/scripts
$ bash ./get-report-specjbb.sh
The script will generate two files: ./jade-ae/results/specjbb/report.csv and ./jade-ae/results/specjbb/rt_curve.csv. These files contain the peak jops numbers and raw data for the response time curve, respectively.
NOTE: Please note that each run of Specjbb2015 will take approximately 2-3 hours. This set of experiments includes testing 21 test points, so it may take up to 63 hours in total. Please make sure to allocate enough time for the experiments.
Run the following script to evaluate baseline and jade.
$ cd ~/jade-ae/scripts
$ bash ./run-hbase.sh
Reports will be automatically generated in the following folder: ~/jade-ae/results/hbase
For instance, you can locate the HBase report for the YCSB mix workload with a 4x heap size (4400m) in the following file ~/jade-ae/results/hbase/4400m-mix.csv. You may have to sort this table by hand in order to reproduce Figure 5.