Schaft? That robot is one bad mother—
This is joint work with Alexandre Ferreira.
Some of us at IBM Research have been doing Docker vs. KVM benchmarking recently and after seeing the discussions of some recent Postgres results I figured I’d give a peek at what we’re seeing from MySQL 5.5. We’re using Docker 1.0 and Ubuntu 13.10 and running on a 2-socket Sandy Bridge E5-2665, giving us 16 total cores.
First let’s look at throughput and latency when running sysbench oltp at different load levels:
There are four different configurations here: MySQL running normally on Linux (native), MySQL under Docker using host networking and a volume (Docker), using a volume but normal Docker networking (Docker NAT), and storing the database within the container filesystem (Docker NAT AUFS). The shape of this curve is what we would expect: throughput increases with load until the machine is saturated, then levels off with a little loss due to contention when overloaded. AUFS introduces significant overhead which isn’t surprising since I/O is going through several layers. NAT also introduces a little overhead, but this isn’t a network-intensive workload.
We also tracked the server’s total CPU utilization during the tests:
We see that CPU utilization is generally proportional to load until the system gets saturated. An interesting twist is that utilization of the three Docker configurations is always lower than running MySQL natively; we don’t know why this is. The real curiosity is shown at the bottom where we graphed the performance difference between Docker and native Linux; this difference is negligible at low load but spikes to 15% around 50 threads and then goes back down to almost zero. MySQL isn’t able to use as much CPU when running under Docker and thus its performance is lower; this looks like mutex contention, but it’s not clear why Linux namespaces would introduce so much overhead on mutexes.
2.5 Gbps never really took off but maybe it will be different if customers are driving.
A great use of a token bucket.
"With the advent of merchant silicon and bare-metal switching, the need to use OpenFlow as technical work-around to a non-technical problem has disappeared."
If this is over 12 months away why is Intel re-pre-announcing it now? Is there nothing else on the roadmap?
Maine is still not allowed on the internet. It’s unfair, but it’s the only way to make sure that Stephen King never has access to a blog.