4 ms·
I think we're starting to see the pendulum swing back towards caring about performance. For years we could just throw more hardware at a problem, and now perfor
by bane 12y ago
I think we're starting to see the pendulum swing back towards caring about performance. For years we could just throw more hardware at a problem, and now performance curves are starting to flatten out. The bandaid has been to use more systems, but old-timers are starting to remember back to their youth, when systems were slow, that there used to be a phase in software development called "performance optimization" that's almost entirely disappeared in the modern discipline.
The amount of personal computing power every person with a modern computer has at their fingertips is staggering and almost nobody has any idea how to use it.
I'd like to think that's starting to change.
- gambiting 12y agoAs someone who has recently obtained a Masters degree in Computing science and who has few friends doing PhDs in big data analysis - I don't see that trend. All the latest research techniques that I see being used currently concentrate at throwing more hardware at the problem if it allows the researches to express their ideas easier - I have a friend who writes algorithms for processing 300GB graphs in Java. And basically the idea is that it really doesn't matter if the algorithm takes 40 hours because 90% of the time is spent garbage collecting. This is still preferable to writing much faster program in C, and spending weeks trying to do memory optimizations and playing around with pointer logic. If there is a need, then Uni will gladly pay for any requested capacity on an Amazon cluster, it's part of the Russel group so they have nearly unlimited budget for research. I am sure that's not the case everywhere, but in my experience researchers gladly trade raw performance for ease of use.
- srean 12y agoTo see that performance matters, talk to the guys who are closer to the electricity bills that the big honking clusters demand. There is quite a bit of money to be saved by using hardware effectively. BTW if 90% time is spent in garbage collection, something is definitely wrong in the choice of tools or methods.
- gambiting 12y agoThat was just an example, I shouldn't have said 90%, sorry.
- bane 12y agoI actually don't disagree with you. I don't really recall seeing much of the nuts and bolts performance work being done when I was in university either. Even back then there was lots of work on distributed computing and such, but almost none on in-system performance optimization. Funny ideas like your program needing to complete before the universe grows cold or shouldn't address more memory than there are particles in the universe don't get lots of attention at the academic level. I think at the engineering level, where the theoretical CS gets turned into real systems, real resource constraints start to become serious problems. Things like the laws of physics and time really become important. So for a 300GB graph, imagine you need to run some computation on data that size every day -- because of business reasons. So now you have to fit it into a 24 hour run-cycle instead of 40. So now you have to figure out how to shave off 40% off the run-time. Boom, you're now having to deal with performance optimizations. For fun, let's say each time I get the result of that data, I can make a decision that earns $10,000, but at 40 hour run-times I can only make that decision 2 or 3 times a week instead of 5. So not only is it 40% too slow, but I'm only making 40% as much money. Also, now you're dedicating a piece of hardware to that one task. And the data better never grow. So you decide to rewrite the entire process in C, spend 2 months doing it, and now it runs in 5 hours instead of 40. I sunk $30k into the rewrite, but I'll make that back inside a week and it has room to grow as the data grows. My other alternative is to figure out some distributed method that will get it down to <24hr run-times, but now I have to manage a compute cluster, with associated facility, electricity and management costs. I probably have to hire, if even part-time, a human or two to manage the cluster, so there goes additional overhead and administrative costs. Maybe I just go with Amazon, but there's downsides with that as well. The solution for this problem is being worked on mostly at the engineering level, not the research one. And that's simply because the business forces are driving it to happen. This is the real-life set of alternative choices computing groups are starting to contend with. It made more sense to distribute when 300GB of graph data was a lot. But now you can get a single machine that can fit that into RAM let alone disk. Single machines are outrageously powerful right now, but nobody has bothered to realize it.
- x0x0 12y ago2 months to rewrite complex distributed systems into relatively more error prone c -- where you are most likely doing lots of things with raw pointers or carefully laying out ram to optimize strides -- seems to me an aggressive schedule. My experience with a somewhat comparable project was perhaps 2 months of dev work, and 1-2 months of bug chasing, with sporadic bug chasing for a while longer.
- x0x0 12y agoI'd bet it's less trading performance for ease-of-use, and more trading performance for the probability of correctness. Which may be just another way of saying ease-of-use.
- angersock 12y agoWe do semi-embedded installs (think teensy datacenters) in medicine. So, on the one hand, I really like the cleanliness of nice service architectures and the HA aspects and whatnot. On the other hand, once installed, that's all the hardware we get to play with for a while, so performance actually matters. It's an interesting space.