Only ~25k RPS when running benchmark via redis-benchmark

[kamyuentse]

I follow the "Get Started" to build&run GarnetServer, then execute a benchmark via redis-benchmark. The result that ~25k RPS on my machine. Did I miss something?

command: redis-benchmark -t set,get -n 10000000 -p 3278 -c 100 -r 10000000

OS: 20.04.1-Ubuntu
dotnet version: 8.0.202
CPU: Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

[badrishc]

1. First, redis-benchmark is known to be extremely heavy on the client side, as it was developed for a different tier of performance.
2. It looks like you have not enabled pipelining which means each request is being fully processed and returned before sending the next in a session. In this mode you are measuring nothing but the performance of localhost network roundtrip time and TCP overhead. Try -P 1024 for example.
3. Either use Resp.benchmark which is carefully designed to ensure client does not become the bottleneck, or if you want a third-party tool then memtier_benchmark (link) is better than redis-benchmark.

[kinthaiofficial]

~25k RPS via redis-benchmark against Garnet is lower than expected given Garnet's claims of multi-million RPS in internal Microsoft benchmarks. A few things to check:

1. Pipeline depth — redis-benchmark defaults to no pipelining, which is latency-bound rather than throughput-bound:

# Without pipelining: tests latency, not throughput
redis-benchmark -h localhost -p 6379 -c 50 -n 100000 GET mykey

# With pipelining: tests actual throughput
redis-benchmark -h localhost -p 6379 -c 50 -n 100000 -P 32 GET mykey

2. Client count — default -c 50 may not saturate Garnet's thread pool:

redis-benchmark -c 500 -P 32 -n 1000000 GET mykey

3. Network path — running benchmark on the same host (loopback) vs. cross-network changes results significantly:

# Same host: should be much higher
# Cross-network: depends on NIC, MTU, etc.

4. Garnet config — make sure thread count is set to match CPU cores:

// garnet.conf
{
  "ThreadPoolMinThreads": 16,
  "ThreadPoolMaxThreads": 64
}

What's your test setup (same machine, cross-network, VM vs. bare metal)? And are you using pipelining? That'll help narrow down whether this is a config issue or a genuine performance ceiling.