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.
As 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.
To 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.
I 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.
2 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.
I'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.
We 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.
Our point in the COST paper was neither that distributed computation systems are bad per se, nor that a laptop is a fine substitute for many of the real-world deployments of distributed systems.
What these posts rail against is a tendency (post-MapReduce/Hadoop) to motivate and evaluate distributed systems solely in terms of their batch performance. This is a particularly common problem in graph processing research, where 2008-era general-purpose systems (i.e. Hadoop) have terrible performance, and so provide a straw-man baseline that a reasonably competent grad student should have had no difficulty in outperforming.
"Does my new system outperform a single thread?" seems like a proper hypothesis to test for a distributed system that claims only a performance motivation. Instead our grad student finds that there is a long list of peer-reviewed systems to compare against, and doing that should be enough, right? It would be fine if those systems had been tested against the first hypothesis, but in most cases they have not, so the new system can be much worse than a single thread (for a small set of benchmarks) and this will go undetected. The COST work is a remedy for this issue.
To return to your original question, I do think that distributed systems are good for some things. First, the OP concerns running graph algorithms on genuinely big data: 128 billion edges, and 3.5 billion vertices. A distributed system should be able to outperform Frank's laptop on that graph, and hopefully this post will encourage developers of new systems to aim higher and evaluate their systems on this scale of data.
The second is that distributed systems offer far more than pure batch performance. For example, geo-distribution is critical to providing a good user experience in web properties; and serving interactive queries with low latency and high availability is a particular interest of mine. We touched on the first part in the Naiad paper [1], and I expect this will become especially important for inference in online learning systems [2]. Hopefully the COST work will force the authors of these systems to try harder in their implementation, and we will all benefit as a result.
My concern was that I saw the blog did the run on real life big(actually big) data and outperform everything on the market. This got me thinking, are we still not there on the problem size/computing power curve so as to justify use of distributed computing?
Of course distributed makes sense in a lot of places specially web and where problem is embarrassingly parallel but running graph algos is a bug chunk of the market and it seems for most firms a laptop will do.
I would wager that an appropriately-sized Spark, GraphLab, or Naiad cluster could outperform a laptop when running PageRank on this graph. Distributed graph processing research got stuck in a rut when the gold standards for evaluation were the Twitter crawl [1] and uk-2007-05 web graph [2]. Now that the Common Crawl has made a much bigger graph available, I hope that we'll see it used in more performance evaluations—with a COST baseline, of course!
I'd also wager that those Spark, GraphLab, or Naiad programs would be far from optimal, by at least an order of magnitude. My ideal would be to see someone attack the distributed graph processing problem with the same eye for systems performance issues that the TritonSort folks brought to the distributed sort problem [3] and MapReduce [4].
When you explain the purpose of COST in this way, it is insightful and the value becomes very apparent. Noting that so many systems have not been compared against a single thread is important, since many developers likely would assume that. Do you feel that your COST work helps provide scalability break points where a developer could choose one system over another when performance is the primary concern? I can see that being helpful in the prototype and early scaling stages, as well as for ad-hoc work.
I think in its current form, COST simply provides a reality check: for some problems, a given system may be completely inappropriate when a simple single-machine implementation exists. Along with raising standards for distributed systems evaluation, I hope that it encourages people to innovate on the single-machine implementations: for example, the GAS idea in PowerGraph [1] is an antecedent of the Hilbert-curve implementation in Frank's first blog post, and it would have been great if the original GAS implementation were separable from the particular distributed system for which it was built.
I fantasize about this work inspiring a competition like the sort benchmark [2], but for single-threaded graph algorithms. Not only would this encourage people to develop usable single-threaded implementations, but it have the side effect of increasing the COST for all distributed implementations....
As for providing scalability break points, I think this work goes some of the way, but doesn't take into account the fact that the input data for a particular real-world computation might not reside in a conveniently-compressed format on a local SSD. Extracting the Common Crawl from an HDFS cluster just to replicate the blog post's experiments could take much longer than the computation itself; extracting the social graph from Facebook's Tao [3] to do offline processing might be similarly non-trivial. Academic experiments don't typically take this cost into account, but recent work like Musketeer [4] has started to consider it, and even offer ways to automate the system choice problem.
Please understand that some of us are not native speakers and try really hard to understand your comments. Particularily in this place, making the effort seems to be obliged, as most of the community produce high-quality comments. What you want to convey with "mmry" is very opaque. I am still not sure of what you mean with it (I believe is "Am I right", but the reference is lost with me, and I don't think I'm the only one).
Well now I feel a little bit embarrased, I was harsh without need. Also, thanks for pointing that out, it's better to know than to be treated with silence!
If you think of the network as part of the memory hierarchy, transferring data to nodes in a distributed system is usually slower than even spinning disk storage. About the only thing slower are USB, floppy drives and tape.
So basically big systems only make sense if the computation time on a single system exceeds the transfer time off of that system to the remote nodes + the time for that data to work it's way up the local node memory hierarchy.
This can happen when the data is big and the computation has unavoidable serialism in each part of the computation. Otherwise you'r likely to be able to coax a single system to just do it faster and with far less complexity than working on a compute cluster.
(I'm ignoring big systems with fast shared memory, like classic super-computers, where there effectively is zero transfer time since each node uses the same memory pool)
> If you think of the network as part of the memory hierarchy, transferring data to nodes in a distributed system is usually slower than even spinning disk storage.
According to Google quick search, the bandwidth of DDR3 is ~12.8 GB/s, or ~104 gbps. As I understand it second-hand (read: cheap) infiniband givies ~40 gpbs, and newer, high-end gives ~100gbps[1]. Now, if you're using regular gigabit ethernet (or even 10gps with cheap-ish switches) -- you'd probably see around 5-8 gbps -- or if we're being very generous, on the order of 1/10th of cpu-ram bandwidth.
Sata2 is rated at ~6 gbps for comparison. In other words, 10gpbs ethernet/infiniband should be on-par with local SSDs, but "proper" infiniband should push the node-node ram-ram performance up compared to that. It appears fiberchannel is also moving towards the ~100 gps mark[2], but isn't there quite yet. And you'd have to have something on the other end -- with high speed networking it is quite natural to assume that both source and destination is RAM.
It's actually pretty crazy with ~40gps infiniband -- I'm considering getting some second-hand gear just to play with at home. Seems both cheaper and more reliable to work with than try to get 10gps ethernet to work (on a home/hobby budget!).
All that said, I think the article makes a lot of great points -- and it does seem a bit funny that people seemingly are thinking of 1 gbps as "fast" when so much has happened on the cpu/ram side since that was remotely true.
Though, you'd be surprised at how often people get stuck on distributed problems where they just can't seem to get their cluster to process more than 80-100MB/sec and it turns out it's because they ran 1Gb ethernet between their nodes.
If I'm reading the numbers right[1], and converting correctly, infiniband is on the order of 5000 - 500 ns. I didn't mean to imply that the network isn't slower, just that it's possible to do quite a lot better than gigabit ethernet quite cheaply.
Yup. Infiniband + RDMA is pretty cool, but RDMA is also rather scary. I mean, writing remotely to another systems (physical) memory. In the same time I do wonder why this is not being used more nowadays.
This is absolutely true! However, I'm not aware of anybody doing this for the graph processing systems being discussed in the post. Nor am I sure what that would mean. Admittedly if something takes 100 times longer to run on multiple machines than it does on a single laptop (as in the previous COST post [1]) then you might care about fault tolerance, but this is a contrived problem if you could achieve the same results faster on yor laptop.
can't help but feel that maybe you missed the point. The point is that an inexpensive single computer, programmed properly, can be seen to blow away gigantic, extremely expensive "big data"-programmed clusters.
if you wanted to run this same algorithm on a 'real' single computer with ecc and high-performance cores, nothing would stop you, and it would get even faster, and still be radically less expensive in every dimension than the "big data" clusters, for this problem set.
Didn't miss the point at all. Lamenting lack of reliable laptops. Well, at least with reliable memory. You can buy a laptop with ECC RAM in the display adapter, but not as main memory! Sigh.
Luckily networking is getting so good that remote desktops feel almost like local most of the time.
So true. If you can't accept the risk of random bit flip in memory, laptops are only good for using a real computer remotely. Memory errors do happen all too often on laptops. Have had it happening in the middle of a AES CBC encrypted data file... How hard would it be to get ECC RAM on laptops for better reliability?
First you should optimize performance locally. By choosing an algorithm and implementation that gives you the best results at data size n you need to deal with most. Lowest O complexity algorithm is not necessarily the best. Worst cases and especially pathological cases also matter.
Then you need to deal with distribution between a large number of systems over network... This will likely force you to choose a different algorithm, but at least you'll have something the benchmark it against from local best case scenario.
When it comes to graphs, I think good partitioning is very important in order to minimize average case latency. So it'd greatly help if you can somehow partition the data to minimize the number of nodes referring to a graph node on a remote system. When walking the graph, instead of immediately sending a query to the target system, it's probably better to batch the queries. Or not if latency counts more. Either way, latency will hurt badly. On a local system it's 100 ns. When the data is on a remote system, you're probably talking about 1000000 ns.
To combat latency, maybe the best way is not to send the request in the first place? Maybe you can ensure nodes referring to a remote system are duplicated to some extent on multiple remote systems. It could help performance -- or hurt it more, like when you need to modify the nodes.
Maybe you should have counters on nodes with connections to nodes on remote systems. If some remote system is referring to it more than local system, migrate the node there instead. It's of course not hard to see how this strategy could blow up. :-)
Please don't take above ideas seriously if you're developing a graph processing system. I wrote those to illustrate the biggest basic problem of distributed systems, latency. Distributed systems are very hard.
This probably isn't crystal clear from this post alone, but for PageRank the 42GB working set has a good mix of read and write. Specifically, about 14GB of target rank vector gets +='d all over the place. If they got hammered randomly, then lots of paging and writing would happen.
I've added the timings from the original post to the tables in this post, which also makes clearer the intended point: the existing systems don't yet have measurements for data at this scale (except Facebook's modified Giraph, which I want to hear more about :))
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.