|
|
|
|
Back to newsletter 200 contents
I was going to go into the concurrency axis from my Devoxx talk (slides) this month, but I've been sidetracked by containerization from a slew of articles I saw (a couple make it into this month's news and tools). I fully understand how beneficial a container can be for providing you with an abstraction away from your environment, after all the JVM is itself a container. But you need to be aware that there is a cost to every level of containerization, and you need to know that the benefits you are gaining outweigh those costs.
When I say container, maybe you think Docker. But we already have multiple levels of containers before you get there. Starting with the hardware, you have the OS overlaying the hardware, so that's the first container. High performance applications often try to bypass the OS: applications that know exactly how they'll operate IO will use a raw partition so that they avoid the filesystem overheads; applications with the tightest latency constraints isolate cores so that they have direct control of the thread running on a core; applications needing to ensure that data transits with no delay lock the process memory into RAM; there are specialist techniques to bypass the OS for the tightest control over network IO. Most applications prefer to use the "OS container" rather than bypass it because it provides a lot of useful features that you don't need to implement yourself.
The next container if you're running on a virtual OS common to most clouds would be the virtualization layer - and many enterprises have to work around it's limitations. Very low latency applications avoid it, instead using dedicated servers; seriously horizontally scaled applications monitor the performance of each virtual server and actually shutdown services when they detect sufficient contention for resources from competing virtual containers using the shared hardware (which means that if you're NOT doing that, your services are the ones that are losing the battle for shared resources).
The JVM layer, whether running on a real or a virtualized OS, is pretty efficient after 20 years of effort in making it efficient primarily by trading more memory for optimal speed, but there are still some things developers find they try to workaround: memory layout at the cache-line level; reducing object copies; etc. (you'll find many techniques in my previous newsletters)
So if you add another container layer like Docker, you should be aware that you are trading whatever features you're getting for some type of performance consideration - that should be an explicit decision, not an "oh dear" after-the-fact one.
Now on to our usual links to articles, tools, news, talks, blogs. And if you need the tips from this month's articles and talks, as ever they are extracted into this month's tips page.
Java performance tuning related news.
Java performance tuning related tools.
Back to newsletter 200 contents
