Redis performance over network

[hi.chr...@gmail.com]

Hi, 

I'm running into performance issues with Redis over network.

My set up: three AWS m3.medium instances, each one running Redis server (one master, two replicated slaves) and sentinel.  I'm not running anything else on these instances.  Transparent huge pages are disabled.  There is no persistence to disk.  My client instance is in the same availability zone and subnet.  The connections between the client and Redis server instances are encrypted with spiped. My keys are short strings and values are strings ranging in length from 0.2KB to 3KB.  I have around 40,000 keys, so not using more than 120MB or so memory total.  There are around 20 connections to Redis at any given time.

Some statistics: latency from client to Redis as measured with `redis-cli --latency` running through spiped averages 0.52ms. Intrinsic latency measured for 100 sec on the Redis master instance averages 0.4042 microseconds but the worst run takes 40,000 longer than average.

The concerning results: GETting a single key with value size 0.2KB takes 0.3ms on the Redis server itself; from client it takes 0.5ms.  However, GETting a single key with value size 3KB takes 0.6ms on Redis server and 77ms from client.  

Could it really be that network is causing this much slowdown?  Are there any basic "gotchas" that I'm missing?  I'm unsure what to think. The network latency, bandwidth, and throughput between the client and Redis instance look fine, so I'm not sure what to think...

Thanks in advance!

Chris

[hva...@gmail.com]

You don't only have the network between your clients and servers.  What are the results without passing the connections through spiped?

[Plhu]

Also,
what is your throughput? How much loaded your vCPUs are? And considering
the fact that you use AWS, don't your VMs suffer from higher CPU steal
rate?

Lukas