Benchmark

This page shows Silly framework's performance in different scenarios.

Test Environment

Test Machine:

• CPU: Intel(R) Core(TM) i7-10700 CPU @ 2.90GHz
• Test Tool: redis-benchmark

Test Command:

./redis-benchmark -t ping -c 100 -n 100000

Test Code

local tcp = require "silly.net.tcp"

local server = tcp.listen {
    addr = "127.0.0.1:6379",
    accept = function(conn)
        while true do
            local l, err = conn:read("\r\n")
            if err then
                break
            end
            if l == "save\r\n" then
                conn:write("*2\r\n$4\r\nsave\r\n$23\r\n3600 1 300 100 60 10000\r\n")
            elseif l == "appendonly\r\n" then
                conn:write("*2\r\n$10\r\nappendonly\r\n$2\r\nno\r\n")
            elseif l == "PING\r\n" then
                conn:write("+PONG\r\n")
            end
        end
    end
}

Test Results

PING_INLINE

====== PING_INLINE ======
  100000 requests completed in 0.42 seconds
  100 parallel clients
  3 bytes payload
  keep alive: 1
  host configuration "save": 3600 1 300 100 60 10000
  host configuration "appendonly": no
  multi-thread: no

Latency by percentile distribution:
0.000% <= 0.087 milliseconds (cumulative count 1)
50.000% <= 0.223 milliseconds (cumulative count 53428)
75.000% <= 0.247 milliseconds (cumulative count 82236)
87.500% <= 0.263 milliseconds (cumulative count 90896)
93.750% <= 0.279 milliseconds (cumulative count 95021)
96.875% <= 0.295 milliseconds (cumulative count 96880)
98.438% <= 0.327 milliseconds (cumulative count 98451)
99.219% <= 0.391 milliseconds (cumulative count 99243)
99.609% <= 0.455 milliseconds (cumulative count 99642)
99.805% <= 0.983 milliseconds (cumulative count 99806)
99.902% <= 1.007 milliseconds (cumulative count 99905)
99.951% <= 1.183 milliseconds (cumulative count 99953)
99.976% <= 1.503 milliseconds (cumulative count 99976)
99.988% <= 1.511 milliseconds (cumulative count 99990)
99.994% <= 1.527 milliseconds (cumulative count 100000)
100.000% <= 1.527 milliseconds (cumulative count 100000)

Cumulative distribution of latencies:
0.006% <= 0.103 milliseconds (cumulative count 6)
25.212% <= 0.207 milliseconds (cumulative count 25212)
97.423% <= 0.303 milliseconds (cumulative count 97423)
99.369% <= 0.407 milliseconds (cumulative count 99369)
99.697% <= 0.503 milliseconds (cumulative count 99697)
99.763% <= 0.607 milliseconds (cumulative count 99763)
99.765% <= 0.703 milliseconds (cumulative count 99765)
99.905% <= 1.007 milliseconds (cumulative count 99905)
99.962% <= 1.207 milliseconds (cumulative count 99962)
99.976% <= 1.503 milliseconds (cumulative count 99976)
100.000% <= 1.607 milliseconds (cumulative count 100000)

Summary:
  throughput summary: 235849.06 requests per second
  latency summary (msec):
          avg       min       p50       p95       p99       max
        0.230     0.080     0.223     0.279     0.367     1.527

Key Metrics:

• Throughput: 235,849 requests/second
• Average Latency: 0.230ms
• P50 Latency: 0.223ms
• P95 Latency: 0.279ms
• P99 Latency: 0.367ms
• Max Latency: 1.527ms

PING_MBULK

====== PING_MBULK ======
  100000 requests completed in 0.44 seconds
  100 parallel clients
  3 bytes payload
  keep alive: 1
  host configuration "save": 3600 1 300 100 60 10000
  host configuration "appendonly": no
  multi-thread: no

Latency by percentile distribution:
0.000% <= 0.143 milliseconds (cumulative count 1)
50.000% <= 0.231 milliseconds (cumulative count 53801)
75.000% <= 0.255 milliseconds (cumulative count 78576)
87.500% <= 0.279 milliseconds (cumulative count 89337)
93.750% <= 0.311 milliseconds (cumulative count 93872)
96.875% <= 0.391 milliseconds (cumulative count 97004)
98.438% <= 0.447 milliseconds (cumulative count 98524)
99.219% <= 0.495 milliseconds (cumulative count 99232)
99.609% <= 0.559 milliseconds (cumulative count 99644)
99.805% <= 0.639 milliseconds (cumulative count 99811)
99.902% <= 0.679 milliseconds (cumulative count 99904)
99.951% <= 0.791 milliseconds (cumulative count 99954)
99.976% <= 0.839 milliseconds (cumulative count 99981)
99.988% <= 0.855 milliseconds (cumulative count 99992)
99.994% <= 0.863 milliseconds (cumulative count 99997)
99.998% <= 0.879 milliseconds (cumulative count 99999)
99.999% <= 0.887 milliseconds (cumulative count 100000)
100.000% <= 0.887 milliseconds (cumulative count 100000)

Cumulative distribution of latencies:
0.000% <= 0.103 milliseconds (cumulative count 0)
20.267% <= 0.207 milliseconds (cumulative count 20267)
93.185% <= 0.303 milliseconds (cumulative count 93185)
97.642% <= 0.407 milliseconds (cumulative count 97642)
99.303% <= 0.503 milliseconds (cumulative count 99303)
99.750% <= 0.607 milliseconds (cumulative count 99750)
99.920% <= 0.703 milliseconds (cumulative count 99920)
99.960% <= 0.807 milliseconds (cumulative count 99960)
100.000% <= 0.903 milliseconds (cumulative count 100000)

Summary:
  throughput summary: 224719.11 requests per second
  latency summary (msec):
          avg       min       p50       p95       p99       max
        0.241     0.136     0.231     0.335     0.479     0.887

Key Metrics:

• Throughput: 224,719 requests/second
• Average Latency: 0.241ms
• P50 Latency: 0.231ms
• P95 Latency: 0.335ms
• P99 Latency: 0.479ms
• Max Latency: 0.887ms

Performance Summary

Test Type	Throughput (req/s)	Avg Latency	P99 Latency
PING_INLINE	235,849	0.230ms	0.367ms
PING_MBULK	224,719	0.241ms	0.479ms

Performance Characteristics

Silly framework exhibits the following performance characteristics:

1. High Throughput - Single process, single thread can reach 200,000+ QPS
2. Low Latency - P99 latency under 0.5ms, average latency around 0.24ms
3. Good Stability - Concentrated latency distribution, max latency controlled within 2ms
4. Resource Efficient - Single-threaded business logic model, no lock contention

Optimization Recommendations

For best performance, it's recommended to:

1. Use Latest Version - Each version includes performance optimizations
2. Enable jemalloc - Compile with make MALLOC=jemalloc
3. CPU Affinity - Use --socket_cpu_affinity and --worker_cpu_affinity to bind CPUs
4. Avoid Blocking Operations - Use async APIs for all I/O
5. Reasonable Coroutine Usage - Avoid creating too many coroutines causing scheduling overhead

Run Your Own Tests

You can reproduce the benchmark with the following commands:

# Compile Silly
make

# Run test server
./silly benchmark.lua

# Run redis-benchmark in another terminal
redis-benchmark -t ping -c 100 -n 100000

Where benchmark.lua contains the test code above.