I've taken the benchmarks from Matthew Rothenberg's phoenix-showdown, updated Phoenix to 0.13.1 and ran the tests on the most powerful machines available at Rackspace.
Framework | Throughput (req/s) | Latency (ms) | Consistency (σ ms) |
---|---|---|---|
Plug | 198328.21 | 0.63ms | 2.22ms |
Phoenix 0.13.1 | 179685.94 | 0.61ms | 1.04ms |
Gin | 176156.41 | 0.65ms | 0.57ms |
Play | 171236.03 | 1.89ms | 14.17ms |
Phoenix 0.9.0-dev | 169030.24 | 0.59ms | 0.30ms |
Express Cluster | 92064.94 | 1.24ms | 1.07ms |
Martini | 32077.24 | 3.35ms | 2.52ms |
Sinatra | 30561.95 | 3.50ms | 2.53ms |
Rails | 11903.48 | 8.50ms | 4.07ms |
Environment | Version |
---|---|
Erlang | 17.5 |
Elixir | 1.1.0-dev |
Ruby | 2.2.0 |
Node | 0.12.4 |
Go | 1.3.3 |
Java | 8u45 |
Rackspace OnMetal IO
CPU Dual 2.8 Ghz, 10 core Intel® Xeon® E5-2680 v2
RAM 128 GB
System Disk 32 GB
Data Disk Dual 1.6 TB PCIe flash cards
Network Redundant 10 Gb / s connections in a high availability bond
Disk I/O Good
Price $2.4658 / hr
Rackspace OnMetal Compute
CPU 2.8 Ghz, 10 core Intel® Xeon® E5-2680 v2
RAM 32 GB
System Disk 32 GB
Network Redundant 10 Gb / s connections in a high availability bond
Disk I/O Good
Price $0.7534 / hr
server:~# MIX_ENV=prod elixir -pa _build/prod/consolidated -S mix phoenix.server
client:~# wrk -t20 -c100 -d30S --timeout 2000 "http://10.184.11.239:4000/showdown"
Running 30s test @ http://10.184.11.239:4000/showdown
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 595.50us 398.08us 62.05ms 99.05%
Req/Sec 8.50k 304.25 9.20k 70.34%
5087667 requests in 30.10s, 10.42GB read
Requests/sec: 169030.24
Transfer/sec: 354.47MB
server:~# MIX_ENV=prod elixir -pa _build/prod/consolidated -S mix phoenix.server
client:~# wrk -t20 -c100 -d30S --timeout 2000 "http://10.184.13.193:4000/showdown"
Running 30s test @ http://10.184.13.193:4000/showdown
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 611.61us 1.04ms 95.57ms 99.05%
Req/Sec 9.04k 656.30 10.98k 71.84%
5408349 requests in 30.10s, 11.08GB read
Requests/sec: 179685.94
Transfer/sec: 376.82MB
server:~# MIX_ENV=prod elixir -pa _build/prod/consolidated -S mix server
client:~# wrk -t20 -c100 -d30S --timeout 2000 "http://10.184.11.239:4000/showdown"
Running 30s test @ http://10.184.11.239:4000/showdown
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 631.83us 2.22ms 217.98ms 99.05%
Req/Sec 9.99k 761.02 18.66k 78.99%
5969613 requests in 30.10s, 12.00GB read
Requests/sec: 198328.21
Transfer/sec: 408.34MB
server:~# ./activator clean stage
server:~# ./target/universal/stage/bin/play-scala -Dhttp.port=5000
client:~# wrk -t20 -c100 -d30S --timeout 2000 "http://10.184.11.239:5000/showdown"
Running 30s test @ http://10.184.11.239:5000/showdown
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 1.89ms 14.17ms 349.18ms 98.57%
Req/Sec 8.80k 1.58k 23.45k 86.04%
5154091 requests in 30.10s, 9.82GB read
Requests/sec: 171236.03
Transfer/sec: 334.12MB
server:~# PUMA_WORKERS=20 MIN_THREADS=1 MAX_THREADS=80 RACK_ENV=production bundle exec puma
client:~# wrk -t20 -c100 -d30S --timeout 2000 "http://10.184.11.239:3000/showdown"
Running 30s test @ http://10.184.11.239:3000/showdown
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 8.50ms 4.07ms 61.71ms 70.94%
Req/Sec 598.10 40.81 740.00 71.03%
357320 requests in 30.02s, 788.91MB read
Requests/sec: 11903.48
Transfer/sec: 26.28MB
server:~# RACK_ENV=production bundle exec puma -t 1:16 -w 20 --preload
client:~# wrk -t20 -c100 -d30S --timeout 2000 "http://10.184.11.239:9292/showdown"
Running 30s test @ http://10.184.11.239:9292/showdown
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 3.50ms 2.53ms 56.23ms 74.62%
Req/Sec 1.54k 414.73 5.33k 65.96%
919897 requests in 30.10s, 1.85GB read
Requests/sec: 30561.95
Transfer/sec: 63.04MB
server:~# GOMAXPROCS=20 GIN_MODE=release go run server.go
client:~# wrk -t20 -c100 -d30S --timeout 2000 "http://10.184.11.239:3000/showdown"
Running 30s test @ http://10.184.11.239:3000/showdown
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 655.59us 573.14us 10.11ms 89.21%
Req/Sec 8.86k 759.03 15.02k 82.84%
5302230 requests in 30.10s, 10.31GB read
Requests/sec: 176156.41
Transfer/sec: 350.61MB
server:~# GOMAXPROCS=20 MARTINI_ENV=production go run server.go
client:~# wrk -t20 -c100 -d30S --timeout 2000 "http://10.184.11.239:3000/showdown"
Running 30s test @ http://10.184.11.239:3000/showdown
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 3.35ms 2.52ms 32.73ms 60.45%
Req/Sec 1.61k 99.91 1.98k 68.77%
962648 requests in 30.01s, 1.87GB read
Requests/sec: 32077.24
Transfer/sec: 63.84MB
server:~# NODE_ENV=production node server.js -w 20
client:~# wrk -t20 -c100 -d30S --timeout 2000 "http://10.184.11.239:3000/showdown"
Running 30s test @ http://10.184.11.239:3000/showdown
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 1.24ms 1.07ms 26.92ms 83.39%
Req/Sec 4.63k 2.37k 13.34k 74.60%
2771185 requests in 30.10s, 5.65GB read
Requests/sec: 92064.94
Transfer/sec: 192.37MB
I've benchmarked ASP.Net, since I do .Net at my day to day job. Rackspace doesn't offer an OnMetal instance with windows, only virtualized instances, so I've done the benchmarks on the most powerful one I could find.
On a Rackspace I/O v1 120GB
instance is listed as:
CPU: 32 vCPUs
RAM: 120 GB
System Disk: 40 GB
Data Disk: 1.2 TB (4 disks)
Network: 10 Gb / s
Disk I/O: Best
Results:
root@client:~# wrk -t20 -c100 -d30S --timeout 2000 "http://10.223.241.211/showdown"
Running 30s test @ http://10.223.241.211/showdown
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 62.98ms 104.13ms 962.77ms 81.84%
Req/Sec 1.02k 435.75 2.92k 70.76%
581748 requests in 30.02s, 1.39GB read
Requests/sec: 19381.05
Transfer/sec: 47.37MB
Results were pretty horrid, but maybe it's the virtualization overhead, or maybe Rackspace's Windows virtualization is bad, or even the 10gbit network between OnMetal instances and virtualized instances is not that good. I might do more testing this weekend to try to figure that out.
It’d be great to see cpu and memory usage alongside.