Efficient data transfer through zero copy in Java | Hacker News Reader

Efficient data transfer through zero copy in Java

ibm.com

53 pointsmixmasteralan12 comments

You might want to read this interesting (albeit slightly outdated) post about leveraging `sendfile` in Node.JS by Peter Griess:

http://blog.std.in/2010/09/09/using-sendfile-with-nodejs/

Is this still best practice? What's the more "up to date" method?

splice and tee are the newer kernel interfaces, sendfile is now implemented using them.

Apparently these primitives haven't found their way to the Node.js land yet... and sendfile is still undocumented.

Not really surprising given the cross platform focus in Node now.

It's a nice idea, but I wonder about the use cases. How often would a Java webserver serve a file from disk without modifying it along the way? If performance is an issue then you'd serve the really static files with nginx (or a similar alternative), and every other file (jsp templates etc.) would be read in memory before being sent on.

(Note that I'm not questioning the technique, I just don't entirely agree on the way the article portrays the benefit of this method in webservers.)

I can think of quite a few use cases: Ebedding a webserver in a Java application. Writing distributed infrastructure software like Hadoop or some kind of messaging system. Anything that caches stuff on a disk to pass it on later. Database systems that replicate transaction logs. Backup and storage systems.

I don't think nginx or web servers in general are the last word ever to be spoken in terms of networking software.

For a specific example I just implemented recently: a database utilizing MappedByteBuffer's to the main DB file that can transferTo the record requested as-is to the requesting client.

The use-cases are certainly specific, but when you need them, it's great.

Often you're serving big chunks of static content with a little bit of dynamic content sprinkled in.

You can effectively concatenate stuff by just writing both to the same channel - which means you can get the zero-copy performance benefits for the static parts of your data if you know exactly how much static data there is before you need to write out something dynamic.

But how could we use this in a servlet container environment? Perhaps we'd wrap the response with a special wrapper to take advantage of this? I'm not sure.

I can see the use if one would recreate a server from the ground up. But that would be a rather large task.

Tomcat has supported zero-copy file transfers for years -- since before this article was written, in fact.

http://tomcat.apache.org/tomcat-6.0-doc/aio.html#Asynchronou...

I agree to a certain extent. If static content is a big issue, you should be using CDNs running barebone server applications anyways.

At the same time, I can imagine plenty of companies dont have the resources or expertise to do these kinds of things.In that case this looks pretty helpful.

Reply on news.ycombinator.com