Project Loom C5M is an experiment to achieve 5 million persistent connections each in client and server Java applications using OpenJDK Project Loom virtual threads.
The C5M name is inspired by the C10k problem proposed in 1999.
The project consists of two simple components, EchoServer and EchoClient.
EchoServer creates many TCP passive server sockets, accepting new connections on each as they come in.
For each active socket created, EchoServer receives bytes in and echoes them back out.
EchoClient initiates many outgoing TCP connections to a range of ports on a single destination server.
For each socket created, EchoClient sends a message to the server, awaits the response, and goes to sleep
for a time before sending again.
EchoClient terminates immediately if any of the follow occurs:
- Connection timeout
- Socket read timeout
- Integrity error with message received
- TCP connection closure
- TCP connection reset
- Any other unexpected I/O condition
After plenty of trial and error, I arrived at the following set of Linux kernel parameter changes to support the target socket scale.
The following article about achieving 1 million persistent connections with Erlang was very helpful in this regard: https://www.metabrew.com/article/a-million-user-comet-application-with-mochiweb-part-1
/etc/sysctl.conf:
fs.file-max=10485760
fs.nr_open=10485760
net.core.somaxconn=16192
net.core.netdev_max_backlog=16192
net.ipv4.tcp_max_syn_backlog=16192
net.ipv4.ip_local_port_range = 1024 65535
net.core.rmem_max = 16777216
net.core.wmem_max = 16777216
net.core.rmem_default = 16777216
net.core.wmem_default = 16777216
net.ipv4.tcp_rmem = 4096 87380 16777216
net.ipv4.tcp_wmem = 4096 87380 16777216
net.ipv4.tcp_mem = 1638400 1638400 1638400
net.netfilter.nf_conntrack_buckets = 1966050
net.netfilter.nf_conntrack_max = 7864200
#EC2 Amazon Linux
#net.core.netdev_max_backlog = 65536
#net.core.optmem_max = 25165824
#net.ipv4.tcp_max_tw_buckets = 1440000
/etc/security/limits.conf:
* soft nofile 8000000
* hard nofile 9000000
Two server-grade hosts were used for bare metal tests:
- HPE ProLiant BL460c Gen10
- 2x 20 core CPU
- 128GB RAM
- Xeon Gold 6230 CPUs
- CentOS 7.9
- Project Loom Early Access Build 19-loom+5-429 (2022/4/4) from https://jdk.java.net/loom/
The server launched with a passive server port range of [9000, 9099].
The client launched with the same server target port range and a connections-per-port count of 50,000, for a total of 5,000,000 target connections.
I stopped the experiment after about 40 minutes of continuous operation. None of the errors, closures, or timeouts outlined above occurred.
The experiment ran for 40 minutes. About 200,000,000 messages were echoed.
Server:
[ebarlas@dct-kvm1.bmi ~]$ ./jdk-19/bin/java --enable-preview -ea -cp project-loom-scale-1.0.0-SNAPSHOT.jar loomtest.EchoServer 0.0.0.0 9000 100 16192 16
Args[host=0.0.0.0, port=9000, portCount=100, backlog=16192, bufferSize=16]
[0] connections=0, messages=0
[1040] connections=0, messages=0
...
[10047] connections=765, messages=765
[11047] connections=7292, messages=7282
...
[2422192] connections=5000000, messages=199113472
[2423193] connections=5000000, messages=199216494
[2424195] connections=5000000, messages=199304727
Client:
[ebarlas@dct-kvm2.bmi ~]$ ./jdk-19/bin/java --enable-preview -ea -cp project-loom-scale-1.0.0-SNAPSHOT.jar loomtest.EchoClient dct-kvm1.bmi.expertcity.com 9000 100 50000 60000 60000 60000
Args[host=dct-kvm1.bmi.expertcity.com, port=9000, portCount=100, numConnections=50000, socketTimeout=60000, warmUp=60000, sleep=60000]
[0] connections=8788, messages=8767
[1013] connections=25438, messages=25438
[2014] connections=45285, messages=45279
[3014] connections=64865, messages=64863
...
[2410230] connections=5000000, messages=199020253
[2411230] connections=5000000, messages=199092940
[2412230] connections=5000000, messages=199188157
The ss command reflects that both client and server had 5,000,000+ sockets open.
Server:
[ebarlas@dct-kvm1.bmi ~]$ ss -s
Total: 5001570 (kernel 0)
TCP: 5000123 (estab 5000010, closed 4, orphaned 0, synrecv 0, timewait 2/0), ports 0
Transport Total IP IPv6
* 0 - -
RAW 0 0 0
UDP 15 11 4
TCP 5000119 17 5000102
INET 5000134 28 5000106
FRAG 0 0 0
Client:
[ebarlas@dct-kvm2.bmi ~]$ ss -s
Total: 5000735 (kernel 0)
TCP: 5000019 (estab 5000006, closed 5, orphaned 0, synrecv 0, timewait 4/0), ports 0
Transport Total IP IPv6
* 0 - -
RAW 0 0 0
UDP 12 8 4
TCP 5000014 13 5000001
INET 5000026 21 5000005
FRAG 0 0 0
The server Java process used 32 GB of committed resident memory and 49 GB of virtual memory. After running for 37m37s, it used 03h39m02s of CPU time.
The client Java process used 26 GB of committed resident memory and 49 GB of virtual memory. After running for 37m12s, it used 06h16m30s of CPU time.
Server:
COMMAND %CPU TIME ELAPSED PID %MEM RSS VSZ
./jdk-19/bin/java --enable-preview -ea -cp project-loom-scale-1.0.0-SNAPSHOT.jar loomtest.EchoServer 582 03:39:02 37:37 35102 24.1 31806188 48537940
Client:
COMMAND %CPU TIME ELAPSED PID %MEM RSS VSZ
./jdk-19/bin/java --enable-preview -ea -cp project-loom-scale-1.0.0-SNAPSHOT.jar loomtest.EchoClient 1012 06:16:30 37:12 19339 20.0 26370892 48553472
A dump of the Java platform threads confirmed expectations. 220 total threads and 80 carrier threads.
As outlined in JEP 425, the number of carrier threads defaults to available processors.
The number of processors reports by the host is 80. 2 CPUs, 20 cores each, hyper-threading (2 threads each).
$ grep ^processor /proc/cpuinfo | wc -l
80
$ ./jdk-19/bin/jstack 28226 > threads.txt
$ grep tid= threads.txt | wc -l
220
$ grep tid= threads.txt | grep ForkJoin | wc -l
80Two EC2 instances were used for cloud tests:
- c5.2xlarge
- 16GB RAM
- 8 vCPU
- Amazon Linux 2 with Linux Kernel 5.10, AMI ami-00f7e5c52c0f43726
- Project Loom Early Access Build 19-loom+5-429 (2022/4/4) from https://jdk.java.net/loom/
The server launched with a passive server port range of [9000, 9049].
The client launched with the same server target port range and a connections-per-port count of 10,000, for a total of 500,000 target connections.
I stopped the experiment after about 35 minutes of continuous operation. None of the errors, closures, or timeouts outlined above occurred.
The experiment ran for 35 minutes. About 17,500,000 messages were echoed.
[ec2-user@ip-10-39-197-143 ~]$ ./jdk-19/bin/java --enable-preview -ea -cp project-loom-scale-1.0.0-SNAPSHOT.jar loomtest.EchoServer 0.0.0.0 9000 50 16192 16
Args[host=0.0.0.0, port=9000, portCount=50, backlog=16192, bufferSize=16]
[0] connections=0, messages=0
[1015] connections=0, messages=0
...
[17020] connections=2412, messages=2412
[18020] connections=10317, messages=10316
...
[2106644] connections=500000, messages=17414982
[2107645] connections=500000, messages=17423544
[ec2-user@ip-10-39-196-215 ~]$ ./jdk-19/bin/java --enable-preview -ea -cp project-loom-scale-1.0.0-SNAPSHOT.jar loomtest.EchoClient 10.39.197.143 9000 50 10000 30000 60000 60000
Args[host=10.39.197.143, port=9000, portCount=50, numConnections=10000, socketTimeout=30000, warmUp=60000, sleep=60000]
[0] connections=131, messages=114
[1014] connections=4949, messages=4949
[2014] connections=13248, messages=13246
...
[2091751] connections=500000, messages=17428105
[2092751] connections=500000, messages=17432046
The ss command reflects that both client and server had 500,000+ sockets open.
Server:
[ec2-user@ip-10-39-197-143 ~]$ ss -s
Total: 500233 (kernel 0)
TCP: 500057 (estab 500002, closed 0, orphaned 0, synrecv 0, timewait 0/0), ports 0
Transport Total IP IPv6
* 0 - -
RAW 0 0 0
UDP 8 4 4
TCP 500057 5 500052
INET 500065 9 500056
FRAG 0 0 0
Client:
[ec2-user@ip-10-39-196-215 ~]$ ss -s
Total: 500183 (kernel 0)
TCP: 500007 (estab 500002, closed 0, orphaned 0, synrecv 0, timewait 0/0), ports 0
Transport Total IP IPv6
* 0 - -
RAW 0 0 0
UDP 8 4 4
TCP 500007 5 500002
INET 500015 9 500006
FRAG 0 0 0
The server Java process used 2.3 GB of committed resident memory and 8.4 GB of virtual memory. After running for 35.12m, it used 14m42s of CPU time.
The client Java process used 2.8 GB of committed resident memory and 8.9 GB of virtual memory. After running for 34.88m, it used 25m19s of CPU time.
Server:
COMMAND %CPU TIME PID %MEM RSS VSZ
./jdk-19/bin/java --enable-preview -ea -cp project-loom-scale-1.0.0-SNAPSHOT.jar loomtest.EchoServer 36.7 00:12:42 18432 14.5 2317180 8434320
Client:
COMMAND %CPU TIME PID %MEM RSS VSZ
./jdk-19/bin/java --enable-preview -ea -cp project-loom-scale-1.0.0-SNAPSHOT.jar loomtest.EchoClient 73.9 00:25:19 18120 17.9 2848356 8901320