JetStream async publish throughput much lower than documentation example (400k vs 130k msgs/sec)

[sereins]

JetStream async publish throughput much lower than documentation example (400k vs 120k msgs/sec)

13:09:34 Starting JetStream asynchronous publisher benchmark [batch=500, clients=1, dedup-window=2m0s, deduplication=false, max-bytes=1,073,741,824, msg-size=128 B, msgs=100,000, multi-subject=false, multi-subject-max=100,000, purge=false, replicas=1, sleep=0s, storage=file, stream=benchstream, subject=jsfoo]
13:09:34 Using stream: benchstream
13:09:34 [1] Starting JetStream asynchronous publisher, publishing 100,000 messages
Finished      0s [================================================================] 100%

NATS JetStream asynchronous publisher stats: 403,828 msgs/sec ~ 49 MiB/sec ~ 2.48us

However, when I run the same benchmark in my environment, I only get:

11:40:20 Starting JetStream asynchronous publisher benchmark [batch=500, clients=1, msg-size=128 B, msgs=100,000, multi-subject=false, multi-subject-max=100,000, purge=false, sleep=0s, stream=benchstream, subject=jsfoo]
11:40:20 Using stream: benchstream
11:40:20 [1] Starting JetStream asynchronous publisher, publishing 100,000 messages
Finished      0s [==========================================================================================================================================] 100%

NATS JetStream asynchronous publisher stats: 134,064 msgs/sec ~ 16 MiB/sec ~ 7.46us

Environment

NATS Server version:
v2.10.14-linux-amd64

NATS CLI version:
0.3

Setup

nats stream rm -f benchstream
nats bench js pub async jsfoo --create

Server Machine Hardware

CPU:   8c  Intel(R) Xeon(R) 6982P-C

Memory:  16G

Disk: nvme0n1  ROTA=0 → NVMe SSD

Network:  
NIC: virtio_net (virtual network interface)
Speed: Unknown (virtio_net does not report real NIC speed)
Duplex: Unknown
Auto-negotiation: off
Port: Other (virtual)
Transceiver: internal
MTU: 8500 (Jumbo frame)

[neilalexander]

v2.10.14 is an old version, please upgrade to v2.12.2 and retry.

[sereins]

Thanks for the reminder! I’ll upgrade to v2.12.2 and retry.

[jnmoyne]

The numbers in the docs are just examples of running the benchmark locally on an M4 MacBook (so nats-server and nats are running on the same machine, over the loopback interface), your mileage will vary and that's why you should run them for yourself, but you should be careful to run in the same environment if you want to get equivalent numbers. For example as soon as you go over a physical network the latency and bandwidth of the network will heavily influence the results you get, same goes for the file storage.