Learning low latency C++ and Java? [closed]

Question

I'm currently in a role where I dont get to write any C++ or Java. However, the role is good because provides me with exposure to the business side (i'm interested in finance).

Eventually I would like to get into high frequency trading infrastructure. Therefore, outside of work hours i'd like to maximise the knowledge I can gain about high performance Java and C++. I already have the Java Performance Tuning book, which is ok but not impressive.

Can people recommend anymore latency blogs/books/websites for learning about making C++/C/Java or even Unix very fast?

Or perhaps making the network parts of the OS (if re-writing Unix components) faster?

EDIT: Or perhaps we could make this THE thread for advice on writing fast code

Answer 1

On the Java side look at the disruptor + Doug Lea's and Brian Goetz's 2 books on concurrent programming.

Answer 2

I'm pretty sure in industry people don't start out writing "fast" code, they start out writing correct and well designed code. Start with something that works. Then measure, via performance tests and benchmarking. Then if performance is good enough stop (and I would wager that if the design is good in java or c++ then you're done here). Otherwise use profiling tools (like JProfiler for java) to look for bottlenecks. Then look for improvements. Over time the code gets "fast".

Aside from that you should probably understand how to write systems that scale horizontally.

Answer 3

I have worked in several organizations completely obsessed with performance. Design intended to be legible, clear, easy to maintain tends to destroy code intended to be fast.

Write the right code, then profile it to be fast.

Here is an article that is helpful, just don't drink too much kool-aid.

What every programmer should know about memory, Part 1

Answer 4

I've to agree with some of your comments, doing optimization in an early stage is not always the best, but that usually depends on the problem you're intending to solve, as the poster mentioned he will create a financial application intended to manage large scale customers and data, then he will need to address the performance issues upfront or he could get fired as soon as the software is released.

To answer the original question, try to create the code, that needs to boost performance, it as simple as possible, even if that means to copy'n paste some code into a large function, sometimes the OO creates little issues in your software that you only noticed when it's running on production. From my experience, bare in mind that one of the greatest killers of the performance is the following instruction:

String a = b + c;

If you're creating a fast performance system, then you need to be aware of this simple, but massive, error and use buffers to avoid memory reallocation. Let me mention an example of this, I'm architect of a solution, in one customer we had a performance problem processing a huge set of data, we did a profile and everything looks great, except for some "native" calls, that we just skipped, after several hours of hard work we decided to tackle does "unrelated" problems, just because we dont have anything left to do on our code, in the top of the list was "StringBuffer.append", then we started to look around in our code to see where the stringbuffer is been used that caused that massive problems and found the following code (which was allover the application)

CacheSystem.add(PREFIX + KEY, obj);

The PREFIX + KEY was the problem... we changed our cache strategy to save a hashtable and use the "key" to access the inner elements, and boom!, that gained 50% of the performance, a method that took 300ms (the first caller of the whole process) changed to 30ms after this change (off course, we'd to change this in several spots).

As someone else mentioned earlier, a system designed to be fast is not as easy to maintain as other kind of software, but it's a trade-off you need to do.