Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/maurice2k/ultrapool

Blazing fast worker pool for Golang
https://github.com/maurice2k/ultrapool

adaptive golang goroutine pool spawning worker

Last synced: 12 days ago
JSON representation

Blazing fast worker pool for Golang

Host: GitHub
URL: https://github.com/maurice2k/ultrapool
Owner: maurice2k
License: mit
Created: 2020-01-16T11:45:22.000Z (almost 5 years ago)
Default Branch: master
Last Pushed: 2023-01-30T10:33:26.000Z (almost 2 years ago)
Last Synced: 2024-08-02T05:12:19.671Z (3 months ago)
Topics: adaptive, golang, goroutine, pool, spawning, worker
Language: Go
Homepage:
Size: 28.3 KB
Stars: 8
Watchers: 4
Forks: 0
Open Issues: 0
Metadata Files:
- Readme: README.md
- License: LICENSE

Awesome Lists containing this project

awesome-blazingly-fast - ultrapool - Blazing fast worker pool for Golang (Go)

README

# ultrapool

*ultrapool* is a blazing fast worker pool for Golang that supports adaptive spawning of new workers while idle workers are being cleaned up after a while.

## Architecture
*ultrapool* is modeled after the worker pool found in [valyala/fasthttp](https://github.com/valyala/fasthttp/blob/master/workerpool.go) but adds some speedups and is not bound to `net.Conn`.

Each worker has it's own channel that are queued in a slice structure whenever a worker is idle. The slice is guarded by a mutex.

While so many channels seem to be a performance issue it turns out that using only a few channels (which are then used by more than one worker) results in much poorer performance.

The main difference to fasthttp's approach is that *ultrapool* uses sharding and a small CAS cache to minimize the locking effects. Locking is done on shard level using a spin lock.

## Example

```golang
wp := ultrapool.NewWorkerPool(func(task ultrapool.Task) {
serveConn(task.(net.Conn))
})

// start the worker pool
wp.Start()

// ...

// add some task
wp.AddTask(conn) // handle connection

// gracefully stop the pool
// Stop() does not wait but also does not terminate any running tasks
wp.Stop()

```

## Benchmarks

Benchmarks have been run on a `Intel(R) Xeon(R) CPU E5-2620 v3 @ 2.40GHz` CPU with 12 cores (+ HT) using Golang 1.18.2.

All worker pools are trying to reduce the overhead of spawning a new worker by re-using spawned workers (time-memory tradeoff).

This is most efficient with work loads that can be processed pretty fast (in the range of a few microseconds). The higher the processing time the less efficient a worker pool gets.

I've benchmarked a few work load scenarios:

### Scenario #1: Just sleep for 1 microsecond

*ultrapool* has a much better performance than spawning go routines.

```
BenchmarkPlainGoRoutines/Sleep_1ms/1-24 2970315 414.0 ns/op 115 B/op 3 allocs/op
BenchmarkPlainGoRoutines/Sleep_1ms/10-24 2422207 489.8 ns/op 123 B/op 3 allocs/op
BenchmarkPlainGoRoutines/Sleep_1ms/50-24 1729738 765.1 ns/op 112 B/op 3 allocs/op
BenchmarkPlainGoRoutines/Sleep_1ms/100-24 1000000 1074 ns/op 126 B/op 3 allocs/op

BenchmarkUltrapoolWorkerpool/Sleep_1ms/1-24 5735512 199.5 ns/op 8 B/op 1 allocs/op
BenchmarkUltrapoolWorkerpool/Sleep_1ms/10-24 5473467 220.5 ns/op 8 B/op 1 allocs/op
BenchmarkUltrapoolWorkerpool/Sleep_1ms/50-24 5488731 225.3 ns/op 9 B/op 1 allocs/op
BenchmarkUltrapoolWorkerpool/Sleep_1ms/100-24 5404957 235.7 ns/op 11 B/op 1 allocs/op

BenchmarkAntsWorkerpool/Sleep_1ms/1-24 840319 1264 ns/op 8 B/op 1 allocs/op
BenchmarkAntsWorkerpool/Sleep_1ms/10-24 904767 1130 ns/op 8 B/op 1 allocs/op
BenchmarkAntsWorkerpool/Sleep_1ms/50-24 1081887 1070 ns/op 9 B/op 1 allocs/op
BenchmarkAntsWorkerpool/Sleep_1ms/100-24 1132861 1053 ns/op 9 B/op 1 allocs/op

BenchmarkEasypoolWorkerpool/Sleep_1ms/1-24 5612767 226.4 ns/op 10 B/op 1 allocs/op
BenchmarkEasypoolWorkerpool/Sleep_1ms/10-24 5639330 214.0 ns/op 17 B/op 1 allocs/op
BenchmarkEasypoolWorkerpool/Sleep_1ms/50-24 5140176 217.8 ns/op 27 B/op 1 allocs/op
BenchmarkEasypoolWorkerpool/Sleep_1ms/100-24 4913846 220.6 ns/op 31 B/op 1 allocs/op

BenchmarkFasthttpWorkerpool/Sleep_1ms/1-24 989583 1242 ns/op 8 B/op 1 allocs/op
BenchmarkFasthttpWorkerpool/Sleep_1ms/10-24 983121 1262 ns/op 8 B/op 1 allocs/op
BenchmarkFasthttpWorkerpool/Sleep_1ms/50-24 928732 1245 ns/op 9 B/op 1 allocs/op
BenchmarkFasthttpWorkerpool/Sleep_1ms/100-24 838867 1279 ns/op 11 B/op 1 allocs/op

BenchmarkGammazeroWorkerpool/Sleep_1ms/1-24 719996 1395 ns/op 24 B/op 2 allocs/op
BenchmarkGammazeroWorkerpool/Sleep_1ms/10-24 829106 1450 ns/op 24 B/op 2 allocs/op
BenchmarkGammazeroWorkerpool/Sleep_1ms/50-24 750172 1417 ns/op 24 B/op 2 allocs/op
BenchmarkGammazeroWorkerpool/Sleep_1ms/100-24 717574 1435 ns/op 24 B/op 2 allocs/op
```

### Scenario #2: Calculate SHA256 over 1024 bytes

*ultrapool* is still ~20% better than spawning go routines.

```
BenchmarkPlainGoRoutines/SHA256_1kB/1-24 2555998 450.3 ns/op 32 B/op 2 allocs/op
BenchmarkPlainGoRoutines/SHA256_1kB/10-24 2323909 464.5 ns/op 32 B/op 2 allocs/op
BenchmarkPlainGoRoutines/SHA256_1kB/50-24 1814096 559.6 ns/op 46 B/op 2 allocs/op
BenchmarkPlainGoRoutines/SHA256_1kB/100-24 1909450 596.9 ns/op 41 B/op 2 allocs/op

BenchmarkUltrapoolWorkerpool/SHA256_1kB/1-24 3005290 397.8 ns/op 8 B/op 1 allocs/op
BenchmarkUltrapoolWorkerpool/SHA256_1kB/10-24 2642110 428.1 ns/op 13 B/op 1 allocs/op
BenchmarkUltrapoolWorkerpool/SHA256_1kB/50-24 2521020 472.0 ns/op 31 B/op 1 allocs/op
BenchmarkUltrapoolWorkerpool/SHA256_1kB/100-24 2216846 467.9 ns/op 36 B/op 1 allocs/op

BenchmarkAntsWorkerpool/SHA256_1kB/1-24 988203 1179 ns/op 8 B/op 1 allocs/op
BenchmarkAntsWorkerpool/SHA256_1kB/10-24 1059818 1080 ns/op 8 B/op 1 allocs/op
BenchmarkAntsWorkerpool/SHA256_1kB/50-24 1081160 1066 ns/op 8 B/op 1 allocs/op
BenchmarkAntsWorkerpool/SHA256_1kB/100-24 1096756 1048 ns/op 9 B/op 1 allocs/op

BenchmarkEasypoolWorkerpool/SHA256_1kB/1-24 1864959 583.3 ns/op 68 B/op 1 allocs/op
BenchmarkEasypoolWorkerpool/SHA256_1kB/10-24 1000000 1691 ns/op 260 B/op 3 allocs/op
BenchmarkEasypoolWorkerpool/SHA256_1kB/50-24 2247285 1427 ns/op 235 B/op 3 allocs/op
BenchmarkEasypoolWorkerpool/SHA256_1kB/100-24 1000000 1255 ns/op 224 B/op 4 allocs/op

BenchmarkFasthttpWorkerpool/SHA256_1kB/1-24 1000000 1182 ns/op 8 B/op 1 allocs/op
BenchmarkFasthttpWorkerpool/SHA256_1kB/10-24 956448 1244 ns/op 8 B/op 1 allocs/op
BenchmarkFasthttpWorkerpool/SHA256_1kB/50-24 858194 1226 ns/op 9 B/op 1 allocs/op
BenchmarkFasthttpWorkerpool/SHA256_1kB/100-24 946658 1212 ns/op 9 B/op 1 allocs/op

BenchmarkGammazeroWorkerpool/SHA256_1kB/1-24 715081 1550 ns/op 24 B/op 2 allocs/op
BenchmarkGammazeroWorkerpool/SHA256_1kB/10-24 745165 1488 ns/op 24 B/op 2 allocs/op
BenchmarkGammazeroWorkerpool/SHA256_1kB/50-24 741728 1465 ns/op 24 B/op 2 allocs/op
BenchmarkGammazeroWorkerpool/SHA256_1kB/100-24 707242 1463 ns/op 24 B/op 2 allocs/op
```

### Scenario #3: Encrypt 1024 byte using AES-128-CBC

Numbers have halved, but *ultrapool* is -- again -- still ~10% better than spawning go routines.

```
BenchmarkPlainGoRoutines/AES_CBC_1kB/1-24 1000000 2184 ns/op 3295 B/op 12 allocs/op
BenchmarkPlainGoRoutines/AES_CBC_1kB/10-24 834814 2591 ns/op 3306 B/op 12 allocs/op
BenchmarkPlainGoRoutines/AES_CBC_1kB/50-24 733308 2506 ns/op 3307 B/op 12 allocs/op
BenchmarkPlainGoRoutines/AES_CBC_1kB/100-24 657081 2448 ns/op 3299 B/op 12 allocs/op

BenchmarkUltrapoolWorkerpool/AES_CBC_1kB/1-24 1000000 1753 ns/op 3279 B/op 11 allocs/op
BenchmarkUltrapoolWorkerpool/AES_CBC_1kB/10-24 859489 2151 ns/op 3301 B/op 11 allocs/op
BenchmarkUltrapoolWorkerpool/AES_CBC_1kB/50-24 850668 7569 ns/op 3343 B/op 11 allocs/op
BenchmarkUltrapoolWorkerpool/AES_CBC_1kB/100-24 832702 2548 ns/op 3353 B/op 11 allocs/op

BenchmarkAntsWorkerpool/AES_CBC_1kB/1-24 1000000 2486 ns/op 3273 B/op 11 allocs/op
BenchmarkAntsWorkerpool/AES_CBC_1kB/10-24 898910 2346 ns/op 3280 B/op 11 allocs/op
BenchmarkAntsWorkerpool/AES_CBC_1kB/50-24 801080 2373 ns/op 3297 B/op 11 allocs/op
BenchmarkAntsWorkerpool/AES_CBC_1kB/100-24 745029 2676 ns/op 3316 B/op 11 allocs/op

BenchmarkEasypoolWorkerpool/AES_CBC_1kB/1-24 1000000 2147 ns/op 3402 B/op 12 allocs/op
BenchmarkEasypoolWorkerpool/AES_CBC_1kB/10-24 1000000 2219 ns/op 3437 B/op 13 allocs/op
BenchmarkEasypoolWorkerpool/AES_CBC_1kB/50-24 1000000 2584 ns/op 3524 B/op 14 allocs/op
BenchmarkEasypoolWorkerpool/AES_CBC_1kB/100-24 1000000 2836 ns/op 3537 B/op 14 allocs/op

BenchmarkFasthttpWorkerpool/AES_CBC_1kB/1-24 1000000 2079 ns/op 3281 B/op 11 allocs/op
BenchmarkFasthttpWorkerpool/AES_CBC_1kB/10-24 851986 1738 ns/op 3279 B/op 11 allocs/op
BenchmarkFasthttpWorkerpool/AES_CBC_1kB/50-24 773848 1904 ns/op 3289 B/op 11 allocs/op
BenchmarkFasthttpWorkerpool/AES_CBC_1kB/100-24 643359 2105 ns/op 3295 B/op 11 allocs/op

BenchmarkGammazeroWorkerpool/AES_CBC_1kB/1-24 669787 2339 ns/op 3288 B/op 12 allocs/op
BenchmarkGammazeroWorkerpool/AES_CBC_1kB/10-24 701192 2103 ns/op 3288 B/op 12 allocs/op
BenchmarkGammazeroWorkerpool/AES_CBC_1kB/50-24 681267 2169 ns/op 3288 B/op 12 allocs/op
BenchmarkGammazeroWorkerpool/AES_CBC_1kB/100-24 698540 2248 ns/op 3288 B/op 12 allocs/op
```

### Scenario #4: Encrypt 8192 bytes using AES-128-CBC

Overall, worker pools are getting less useful; *ultrapool* is still ~4% better than spawning go routines.

```
BenchmarkPlainGoRoutines/AES_CBC_8kB/1-24 1000000 7656 ns/op 20938 B/op 12 allocs/op
BenchmarkPlainGoRoutines/AES_CBC_8kB/10-24 453580 11320 ns/op 20984 B/op 12 allocs/op
BenchmarkPlainGoRoutines/AES_CBC_8kB/50-24 199285 12905 ns/op 21491 B/op 12 allocs/op
BenchmarkPlainGoRoutines/AES_CBC_8kB/100-24 113212 9477 ns/op 20273 B/op 12 allocs/op

BenchmarkUltrapoolWorkerpool/AES_CBC_8kB/1-24 497756 6773 ns/op 20950 B/op 11 allocs/op
BenchmarkUltrapoolWorkerpool/AES_CBC_8kB/10-24 486513 7189 ns/op 20971 B/op 11 allocs/op
BenchmarkUltrapoolWorkerpool/AES_CBC_8kB/50-24 444885 9082 ns/op 21037 B/op 12 allocs/op
BenchmarkUltrapoolWorkerpool/AES_CBC_8kB/100-24 414063 10335 ns/op 21054 B/op 12 allocs/op

BenchmarkAntsWorkerpool/AES_CBC_8kB/1-24 99210 11919 ns/op 20951 B/op 11 allocs/op
BenchmarkAntsWorkerpool/AES_CBC_8kB/10-24 99079 11229 ns/op 20981 B/op 11 allocs/op
BenchmarkAntsWorkerpool/AES_CBC_8kB/50-24 93127 11519 ns/op 21048 B/op 12 allocs/op
BenchmarkAntsWorkerpool/AES_CBC_8kB/100-24 78806 12913 ns/op 21079 B/op 13 allocs/op

BenchmarkEasypoolWorkerpool/AES_CBC_8kB/1-24 252812 7480 ns/op 21043 B/op 12 allocs/op
BenchmarkEasypoolWorkerpool/AES_CBC_8kB/10-24 434630 11067 ns/op 21115 B/op 13 allocs/op
BenchmarkEasypoolWorkerpool/AES_CBC_8kB/50-24 443392 10109 ns/op 21106 B/op 13 allocs/op
BenchmarkEasypoolWorkerpool/AES_CBC_8kB/100-24 389287 10792 ns/op 21213 B/op 13 allocs/op

BenchmarkFasthttpWorkerpool/AES_CBC_8kB/1-24 414046 8100 ns/op 20966 B/op 11 allocs/op
BenchmarkFasthttpWorkerpool/AES_CBC_8kB/10-24 378027 8116 ns/op 20983 B/op 11 allocs/op
BenchmarkFasthttpWorkerpool/AES_CBC_8kB/50-24 377748 8923 ns/op 20989 B/op 11 allocs/op
BenchmarkFasthttpWorkerpool/AES_CBC_8kB/100-24 356851 10058 ns/op 20997 B/op 11 allocs/op

BenchmarkGammazeroWorkerpool/AES_CBC_8kB/1-24 497506 7220 ns/op 20952 B/op 12 allocs/op
BenchmarkGammazeroWorkerpool/AES_CBC_8kB/10-24 474954 6491 ns/op 20952 B/op 12 allocs/op
BenchmarkGammazeroWorkerpool/AES_CBC_8kB/50-24 458689 6995 ns/op 20952 B/op 12 allocs/op
BenchmarkGammazeroWorkerpool/AES_CBC_8kB/100-24 422271 7307 ns/op 20953 B/op 12 allocs/op
```

## License

*ultrapool* is available under the MIT [license](LICENSE).