Benchmarks

Redistill is tuned for single-instance performance, achieving up to 9.07M operations per second with sub-0.5 ms p50 latency, while remaining fully compatible with Redis clients.

Throughput (c7i.16xlarge)
9.07M ops/s
4.5× Redis, 1.7× Dragonfly
Latency p50
0.479 ms
5.0× faster than Redis
Bandwidth
1.58 GB/s
4.7× Redis, 1.7× Dragonfly

Quick summary

On an AWS c7i.16xlarge instance (64 vCPUs, 128 GB RAM), Redistill outperforms both Redis and Dragonfly in a 1:1 mixed SET:GET workload using memtier_benchmark.

Metric Redistill Dragonfly Redis vs Redis vs Dragonfly
Throughput 9.07M ops/s 5.43M ops/s 2.03M ops/s 4.5× 1.7×
Bandwidth 1.58 GB/s 923 MB/s 337 MB/s 4.7× 1.7×
Avg latency 0.524 ms 0.877 ms 2.000 ms 3.8× faster 1.7× faster
p50 latency 0.479 ms 0.807 ms 2.383 ms 5.0× faster 1.7× faster
p99 latency 1.215 ms 1.975 ms 2.959 ms 2.4× faster 1.2× faster
p99.9 latency 1.591 ms 2.559 ms 4.159 ms 2.6× faster 1.6× faster

Redistill processed 544M total operations in this test (2.7× more than Dragonfly, 4.5× more than Redis) with consistently lower tail latency and no errors across all systems.

Throughput, latency and bandwidth

The upstream documentation includes simple visualizations of the competitive benchmark. To reuse those charts, copy the images from the main project docs/img/ directory into this site's assets/img/ directory: throughput.png, latency.png, and bandwidth.png.

Throughput comparison (higher is better)

Throughput comparison: Redistill vs Dragonfly vs Redis

Latency p50 (lower is better)

Latency p50 comparison: Redistill vs Dragonfly vs Redis

Bandwidth comparison (higher is better)

Bandwidth comparison: Redistill vs Dragonfly vs Redis

Detailed results (c7i.8xlarge)

On a slightly smaller AWS c7i.8xlarge (32 vCPUs, 64 GB RAM) using redis-benchmark, Redistill matches or beats Redis across all workloads, with especially strong gains on read-heavy tests.

Workload Redistill Redis Improvement
Interactive (-c 1, -P 1) 39K ops/s 39K ops/s Similar
Production SET (-c 50, -P 16) 2.08M ops/s 1.69M ops/s +23%
Production GET (-c 50, -P 16) 2.05M ops/s 1.94M ops/s +6%
High concurrency SET (-c 300, -P 32) 2.67M ops/s 2.23M ops/s +20%
High concurrency GET (-c 300, -P 32) 3.47M ops/s 2.71M ops/s +28%
Extreme SET (-c 100, -P 64) 2.72M ops/s 2.58M ops/s +5%
Extreme GET (-c 100, -P 64) 5.32M ops/s 3.06M ops/s +74%
Ultra SET (-c 500, -P 128) 2.74M ops/s 2.74M ops/s Equal
Ultra GET (-c 500, -P 128) 6.87M ops/s 3.47M ops/s +98%

Redistill was tuned with 2048 shards, batch size 256, and buffer pool size 2048 – the recommended configuration for balanced workloads described in the Performance Tuning Guide.

Benchmark methodology

On this page

Hardware

  • AWS c7i.16xlarge – Intel Xeon (4th Gen Sapphire Rapids), 64 vCPUs, 128 GB RAM, up to 50 Gbps network
  • AWS c7i.8xlarge – Intel Xeon (4th Gen Sapphire Rapids), 32 vCPUs, 64 GB RAM, up to 12.5 Gbps network

Software

  • Redis 7.0.x (latest stable)
  • Dragonfly (latest stable)
  • Redistill v1.1.2
  • memtier_benchmark and redis-benchmark from Redis 7

memtier_benchmark configuration (c7i.16xlarge)

memtier_benchmark \
  --server=localhost \
  --port=6379 \
  --protocol=redis \
  --threads=8 \
  --clients=20 \
  --pipeline=30 \
  --data-size=256 \
  --ratio=1:1 \
  --test-time=60 \
  --key-pattern=R:R

redis-benchmark configuration (c7i.8xlarge)

# Production workload example
redis-benchmark -t set,get -n 1000000 -c 50 -P 16 -q

# Extreme workload example
redis-benchmark -t set,get -n 1000000 -c 100 -P 64 -q

# Ultra workload example
redis-benchmark -t set,get -n 1000000 -c 500 -P 128 -q

All systems were configured as close as possible, with persistence disabled for a fair comparison with Redistill's default (no persistence). Full raw results are available in the upstream repository under tests/benchmarks/benchmark_results_memtier/.