Follow Slashdot stories on Twitter

 



Forgot your password?
typodupeerror
×
Communications Databases Social Networks Stats Twitter

Harvard/MIT Student Creates GPU Database, Hacker-Style 135

First time accepted submitter IamIanB writes "Harvard Middle Eastern Studies student Todd Mostak's first tangle with big data didn't go well; trying to process and map 40 million geolocated tweets from the Arab Spring uprising took days. So while taking a database course across town at MIT, he developed a massively parallel database that uses GeForce Titan GPUs to do the data processing. The system sees 70x performance increases over CPU-based systems, and can out crunch a 1000 node MapReduce cluster, in some cases. All for around $5,000 worth of hardware. Mostak plans to release the system under an open source license; you can play with a data set of 125 million tweets hosted at Harvard's WorldMap and see the millisecond response time." I seem to recall a dedicated database query processor that worked by having a few hundred really small processors that was integrated with INGRES in the '80s.
This discussion has been archived. No new comments can be posted.

Harvard/MIT Student Creates GPU Database, Hacker-Style

Comments Filter:
  • by Anonymous Coward on Monday April 22, 2013 @06:23PM (#43520555)

    1. Facebook would like to have a discussion with him.
    2. The FBI would like to have a discussion with him.

  • by Anonymous Coward on Monday April 22, 2013 @06:26PM (#43520573)

    I want to know why GPUs are so much better at some tasks than CPUs? And, why aren't they used more often if they are orders of magnitude faster?

    Thanks.

    • by Anonymous Coward on Monday April 22, 2013 @06:29PM (#43520597)

      Sprinters can run really fast. So, if speed is important in other sports, why aren't the other sports full of sprinters? Because being good at one thing doesn't mean you're well-suited to do everything. A sprinter who can't throw a ball is going to be terrible at a lot of sports.

    • by gubon13 ( 2695335 ) on Monday April 22, 2013 @06:36PM (#43520659)
      Sort of a lazy effort on my part to not summarize, but here's a great explanation: https://en.bitcoin.it/wiki/Why_a_GPU_mines_faster_than_a_CPU [bitcoin.it].
    • by PhamNguyen ( 2695929 ) on Monday April 22, 2013 @06:40PM (#43520687)

      GPUs are much faster for code that can be parallelized (basically this means having many cores doing the same thing, but on different data). However there is a signficant complexity in isolating hte parts of the code that can be done in parallel. Additionally, there is a cost to moving data to the GPU's memory, and also from the GPU memory to the GPU cores. CPU's on the other hand, have a cache architecture that means that much of the time, memory access is extremely fast.

      Given progress in the last 10 years, the set of algorithms that can be parallelized is very large. So the GPU advantage should be overwhelming. The main issue is that the complexity writing a program that does things on the GPU is much higher.

      • by gatkinso ( 15975 ) on Monday April 22, 2013 @07:04PM (#43520885)

        >> The main issue is that the complexity writing a program that does things on the GPU is much higher.

        Not so much. There is programming overhead, but it isn't too bad.

      • by Morpf ( 2683099 ) on Monday April 22, 2013 @07:19PM (#43520965)

        Close, but not quite correct.

        The point is GPUs are fast doing the same operation on multiple data. (e.g. multiplying a vector with a scalar) The emphasize is on _same operation_, which might not be the case for every problem one can solve parallel. You will loose speed as soon your elements of a wavefront (e.g. 16 threads, executed in lockstep) diverge into multiple execution paths. This happens if you have something like an "if" in your code and one for one work item the condition is evaluated to true and for another it's evaluated to false. Your wavefront will only be executed one path at a time, so your code becomes kind of "sequential" at this point. You will loose speed, too, if the way you access your GPU memory does not fulfill some restrictions. And by the way: I'm not speaking about some mere 1% performance loss but quite a number. ;) So generally speaking: not every problem one can solve in parallel can be efficiently solved by a GPU.

        There is something similar to caches in OpenCL: it's called local data storage, but it's the programmers job to use them efficiently. Memory access is always slow if it's not registers you are accessing, be it CPU or GPU. When using a GPU you can hide part of the memory latency by scheduling way more threads than you can physically run and always switch to those who aren't waiting for memory. This way you waste less cycles waiting for memory.

        I support your view writing for GPU takes quite a bit of effort. ;)

        • by loneDreamer ( 1502073 ) on Monday April 22, 2013 @10:35PM (#43521931)
          True. Nevertheless, using it for databases when data is cached seems like a neat idea. Lots of "_same operation_" for let's say, selecting all tuples with a specific value on a huge table.
        • by PhamNguyen ( 2695929 ) on Tuesday April 23, 2013 @01:30AM (#43522575)
          I was intentionally simplifying, but I agree with your more detailed exposition. I did understate the extent to which fundamental issues related to the GPU architecture are still relevant. My own experience is in embarrassingly parallelizable problems so my knowledge of these issues is nor very deep.
        • by dj245 ( 732906 ) on Tuesday April 23, 2013 @11:09AM (#43525867)

          Close, but not quite correct.

          The point is GPUs are fast doing the same operation on multiple data. (e.g. multiplying a vector with a scalar) The emphasize is on _same operation_, which might not be the case for every problem one can solve parallel. You will loose speed as soon your elements of a wavefront (e.g. 16 threads, executed in lockstep) diverge into multiple execution paths. This happens if you have something like an "if" in your code and one for one work item the condition is evaluated to true and for another it's evaluated to false. Your wavefront will only be executed one path at a time, so your code becomes kind of "sequential" at this point. You will loose speed, too, if the way you access your GPU memory does not fulfill some restrictions.

          I'm not an expert on this subject, but with hundreds, or even thousands, of GPU cores, why not just run the calculation for all cases of an if/then and then toss out the ones that don't pan out? It's not a very efficient way to do things, but it could work.

          I believe this is the principle of quantum computing. Process all possible answers simultaneously and then pick the right one(s).

          • by Morpf ( 2683099 ) on Tuesday April 23, 2013 @11:46AM (#43526405)

            Actually this is what is done on GPUs. Think of it this way: You have a number of "processors" which share one control flow. The number of "processors" sharing one control flow on a AMD 79xx is 64. Now all these "processors" evaluate the if-statement. If it's true for some and false for other "processors" than both paths are executed sequentially. Those "processors" which would normally not run -because they belong to the other branch- are masked, so they don't execute the instructions. If the if-statement is evaluated true or false for all "processors" in this group only one path is taken.

            Note: Actually there are 16 processors working parallel but they execute 4 times, each time on a different work item. So the correct word for "processor" would be work item.

      • by BitZtream ( 692029 ) on Monday April 22, 2013 @09:23PM (#43521589)

        Parallelization is not why GPUs are fast, its a side effect of rendering pixels, nothing more.

        GPUs are fast because they do an extremely limited number of things REALLY REALLY fast, and when you're doing graphics ... well guess what, its all pretty much doing those few things the GPU does well over and over again, per pixel (or vertex). They are parallelized because those simply, super fast processors are also small from a chip perspective, so stuffing a ton of them on the chip so it can do many pixels in parallel works, again, because all those pixels get treated the same way with a very limited number of well known operations performed on them.

        They are not replacing CPUs because something like a simple if statement doesn't pause one processor, it pauses them ALL, and then top it off with the GPU being absolutely horrible (speed wise) at dealing with an IF statement. In shaders, you can get by with an if on a uniform because it only has to be calculated once and a decent driver can optimize the if away early on be for sending it to the cell processors on the CPU. Do IFs on an attribute (say a vertex or texture coord) and watch your GPU crawl like a snail.

        Parallelization in GPUs is a direct result of the fact that they perform the same task on massive arrays of data. Since the code works on individual cells in the array individually, there is no 'race' condition possibility in the code, so its ready to run concurrently. Adding a new shader cell effectively gives you more speed without any sort of programmer effort what so ever.

        The reason these parallel cells can work together so fast is also because the silicon works in lock step. (thats why IFs or attributes kill performance). Basically each line of the shader program executes side by side on all the shader cells at once. This makes all sorts of neat silicon based performance tricks possible.

        Where you get screwed however is those IFs (All branching instructions really) because if any one shader cell has to run a branch of code, they ALL run the code, and then just discard the results. So when you write branching code in a shader, you are almost certainly going to run every code path provided if you use the wrong data for your branch.

        • by PhamNguyen ( 2695929 ) on Tuesday April 23, 2013 @01:21AM (#43522551)
          What you are describing is GPU computing 5 to 10 years ago. Now, (1) you don't wrote shaders you write kernels. (2) a GPU can do most of the functions of a CPU, the difference is in things like branch prediction and caching. (3) threads execute in blocks of 16 or some other round number. There is no performance loss as long as all threads in the same block take the same execution path.
        • by Morpf ( 2683099 ) on Tuesday April 23, 2013 @12:08PM (#43526711)

          Actually parallelization IS why GPUs are fast. You have some restrictions but it's the parallel execution which gives you the boost in performance.

          The things a GPU can do are not so limited as you might think.

          The statement about "if" pausing all processors is wrong. On my card 64 work items are executed in lockstep on 16 processors in something called a wavefront. Now I have way more processors on the card. Furthermore only when the if statement in the control flow is evaluated true for some work items and false for others you get hit by degraded performance, as you then need to execute both paths sequentially. If the if statement is evaluated the same for the whole wave front you don't loose anything. This works on attributes of the work items or other run time data, too.

          There is a good possibility for race conditions in OpenCL code. Depending on the algorithms and optimizations one has to synchronize or use different places for input and output.

    • by gatkinso ( 15975 ) on Monday April 22, 2013 @07:01PM (#43520857)

      This is a gross simplification, glossing over the details and not correct in some aspects... but close enough.

      SIMD - single instruction multiple data. If you have thousands or millions of elements/records/whatever that all require the exact same processing (gee, say like a bunch of polygons being rotated x radians perhaps????) then this data can all be arranged into a bitmap and loaded onto the GPU at once. The GPU then performs the same operation on your data elements simultaneously (simplification). You then yank off the resultant bitmap and off you go. CPU arranges data, loads and unloads the data. GPU crunches it.

      A CPU would have to operate on each of these elements serially.

      Think of it this way - you are making pennies. GPU takes a big sheet of copper and stamps out 10000 pennies at a time. CPU takes a ribbon of copper and stamps out 1 penny at a time... but each iteration of the CPU is much faster than each iteration of the GPU. Perhaps the CPU can perform 7000 cycles per second, but the GPU can only perform 1 cycle per second. At the end of that second... the GPU produced 3000 more pennies than the CPU.

      Some problem sets are not SIMD in nature. Lot's of branhcing or relienace on the value of neighboring elements. This will slow the GPU processing down insanely. FPGA is far better (and more expensive, and more difficult to program) than GPU for this. CPU is better as well.

      • by crutchy ( 1949900 ) on Monday April 22, 2013 @07:42PM (#43521085)

        not sure if i'm right, but i tend to think of any gpu-based application as having to construct data like pixels on a screen or image (since that's what gpu's are primarily designed to handle)

        a cpu treats each pixel separately, whereas a gpu can process multiple pixels simultaneously

        problem comes about if you try to feed data into a gpu that isn't like pixels

        is the programming difficulties in trying to trick the gpu into thinking it's processing pixels even though it may be processing bitcoin algorithms etc?

        • by Morpf ( 2683099 ) on Monday April 22, 2013 @08:05PM (#43521213)

          Well, you don't have to trick the GPU in thinking it processes pixels. You can do general purpose computation with a language quite similar to C99.

          You are right in that way, that you partition your problem in many subelements. In OpenCL those are called work items. But those are more like identical threads than pixels. Sometimes one maps the work items on a 2d or 3d grid if the problem domain fits. (e.g. image manipulation, physics simulation)

          Actually it's not that hard implementing "normal" algorithms on a GPU. For example the bitcoin mining algorithm can be implemented quite straight forward. It may even look almost the same as a C method programmed for a CPU. The programming is a bit difficult as you have many restrictions to obey to get good performance out of a GPU.

    • by UnknownSoldier ( 67820 ) on Monday April 22, 2013 @07:24PM (#43520991)

      If one woman can have a baby in 9 months, then 9 women can have a baby in one month, right?

      No.

      Not every task can be run in parallel.

      Now however if your data is _independent_ then you can distribute the work out to each core. Let's say you want to search 2000 objects for some matching value. On a 8-core CPU you would need 2000/8 = 250 searches. On the Titan each core could process 1 object.

      There are also latency vs bandwidth issues, meaning it takes time to transfer the data from RAM to the GPU, process, and transfer the results back, but if the GPU's processing time is vastly less then the CPU, you can still have HUGE wins.

      There are also SIMD / MIMD paradigms which I won't get into, but basically in layman's terms means the SIMD is able to process more data in the same amount of time.

      You may be interested in reading:
      http://perilsofparallel.blogspot.com/2008/09/larrabee-vs-nvidia-mimd-vs-simd.html [blogspot.com]
      http://stackoverflow.com/questions/7091958/cpu-vs-gpu-when-cpu-is-better [stackoverflow.com]

      When your problem domain & data are able to be run in parallel then GPU's totally kick a CPU's in terms of processing power AND in price. i.e.
      An i7 3770K costs around $330. Price/Core is $330/8 = $41.25/core
      A GTX Titan costs around $1000. Price/Core is $1000/2688 = $0.37/core

      Remember computing is about 2 extremes:

      Slow & Flexible < - - - > Fast & Rigid
      CPU (flexible) vs GPU (rigid)

      * http://www.newegg.com/Product/Product.aspx?Item=N82E16819116501 [newegg.com]
      * http://www.newegg.com/Product/Product.aspx?Item=N82E16814130897 [newegg.com]

      • by Anonymous Coward on Monday April 22, 2013 @10:04PM (#43521785)

        Now however if your data is _independent_ then you can distribute the work out to each core.

        Let me translate this into a woman-baby analogy: if one woman can have a baby in 9 months, then 9 women can have 9 babies in 9 months. At first the challenge is joggling with the timing of dates and dividing the calendar for conception events as near as possible to each other to keep up the efficiency and synchronization. Afterwards the challenge is the alimony, paying up college and particularly the Thanksgiving, when the fruits of the labor come together.

      • If one woman can have a baby in 9 months, then 9 women can have a baby in one month, right?

        No.

        You're wrong, otherwise we'd need close to 130 million months per year. Furthermore, the 9 women have their 9 babies after ~9 months yielding in an average production rate of 1bpm (one baby per month) from this group of women -- If kept perpetually pregnant. If we put 90 women in the baby farm they will produce TEN Babies Per Month.

        Some people's kids, I swear -- They must have botch the batch of logic circuits in your revision; This is Matrixology 101.

      • by raxhelm ( 2905073 ) on Tuesday April 23, 2013 @08:14AM (#43524037)

        "When your problem domain & data are able to be run in parallel then GPU's totally kick a CPU's in terms of processing power AND in price. i.e.
        An i7 3770K costs around $330. Price/Core is $330/8 = $41.25/core
        A GTX Titan costs around $1000. Price/Core is $1000/2688 = $0.37/core"

        That's a very unfair comparison. For one the i7-3770k has SIMD as well (8-wide AVX). A better comparison is maximum GFLOPs/s.
        The max for the 3700k is frequency * 8 (AVX) * 2(simultaneous multiplication and addition) * 4 cores. Let's assume the frequency is 4.0 GHz (base turbo is 3.8 - mine is overclocked to 4.4). That's 256 SP (single precision) GFLOPs/s and 128 DP (double precision) GFLOPs/s. According to wikipedia the Titan can do 4500 SP GFLOPs/s and 1300 DP GFLOPs/s. That's 19x as many SP GFLOP/s and 10x as many DP GFLOPs. In USD that's 6x as many SP GFLOPs/s per USD and 3x as many DP GFLOPs/s per USD.

        When Haswell comes out in June the max will double due to FMA3. That means the Titan will deliver 3x SP GFLOPs/s per USD and only 1.5x DP GLOPs/s per USD. Hawell's DP performance and price will already be better than the GK 104 chip (GTX 680). However, AMD Radeon will still have an advantage in performance and performance/price both for SP and DP.

        Likely the student who made his/her database on the GPU did not optimize the CPU code because if he/she did the advantage would have been a lot less that 70x.

      • by Bobfrankly1 ( 1043848 ) on Tuesday April 23, 2013 @09:32AM (#43524691)

        If one woman can have a baby in 9 months, then 9 women can have a baby in one month, right?...

        I'm sorry, this is slashdot. You must include an obligatory car analogy to get your point across.

    • by BitZtream ( 692029 ) on Monday April 22, 2013 @09:08PM (#43521521)

      They do ONE thing well. Floating point ops. EVERYTHING ELSE THEY SUCK AT, including simple logic checks, like if statements are painfully mind numbingly slow on the GPU.

    • by Anonymous Coward on Monday April 22, 2013 @09:59PM (#43521763)

      They're massively more parallel, running many more smaller simpler cores.

      It's the same reason these guys can make a 16 core parallel computer for $99.... the cores are focused on their job so they can be smaller and cheaper and can put more on a die.
      http://www.kickstarter.com/projects/adapteva/parallella-a-supercomputer-for-everyone/

      So these guys can run 8 daughter boards, with 64 cores per board, 512 cores, and it looks like they plan on scaling to 4096 cores because they use the top 12 bits of the address as the core routing id.
      The tradeoff with all those cores is they're dirt simple cores, moves, adds, branches, and some floating point ops (misses divide even, its done in software, but then for signal processing and multiply-add is the one that needs to be fast and its coded as a single instruction).

      If you read up on your high end graphics card it might have 900+ CUDA cores, really just ALU cores, the actual thread running cores are far fewer than that. But the ALU's can be run in parallel.

      So a vector multiply is done as a parallel operation on these ALU blocks, and many other operations break down to be parallel in the same way.

    • by H0p313ss ( 811249 ) on Monday April 22, 2013 @11:17PM (#43522119)

      I want to know why GPUs are so much better at some tasks than CPUs? And, why aren't they used more often if they are orders of magnitude faster?

      Thanks.

      I'm glad you put the preface in there, because it's basic comp. sci.

    • by r2kordmaa ( 1163933 ) on Tuesday April 23, 2013 @12:06AM (#43522273)
      CPU has small number of very complex cores, good for fast decision making, eg managing opsys resources GPU has lots of very simple cores, useless for decision making, but great for parallel number crunching
    • by sapgau ( 413511 ) on Tuesday April 23, 2013 @05:28PM (#43530593) Journal

      Part of the answer is the "magic" of matrix math. You can represent multiple linear equations in every row of a matrix and when you apply one operation (add, multiply, etc) you performed it on all your encoded equations inside the matrix.

      If you can, for example, represent your problem in a linear equation (algebra) then you can also formulate 50 similar equations. You want to "transform" all your equations by some operation (lets say divide by 20), so instead of calculating 50 times that operation for every equation you just do it once if you encode them in a matrix.

      This started in graphics when you wanted to apply the same operation to many pixels on a screen (i.e. apply a shading rule), so that's why video cards have these massive processing power on pixels.

      https://en.wikipedia.org/wiki/Matrix_(mathematics) [wikipedia.org]

  • Slashdotted? I happened to catch the story just as it went live, and hit the link to the service. After scrolling the map and getting a couple of updates: Database is down for maintenance. The front end may not be as high performance as the back... or it may have been coincidence.

  • by roman_mir ( 125474 ) on Monday April 22, 2013 @06:33PM (#43520627) Homepage Journal

    as the TFS states he uses GPUs to do the data processing, but you are never going to believe what he uses to store the actual data, you won't believe it, that's why it's not mentioned in TFS. Sure sure, it's PostgreSQL, but the way the data was stored physically was in the computer monitor itself. Yes, he punched holes in computer monitors with a chisel and used punch card readers to read those holes from the screens.

    • by eyenot ( 102141 ) <eyenot@hotmail.com> on Monday April 22, 2013 @06:45PM (#43520737) Homepage

      Mod parent up!

      Also: I heard he's using the printer port for commuication. By spooling tractor feed paper between two printers in a loop, and by stopping and starting simultaneous paper-feed jobs, he can create a cybernetic feedback between the two printers that results in a series of quickly occurring "error - paper jam" messages that (due to two taped-down "reset" buttons) are quickly translated from the wide bandwidth analog physical matrix into kajamabits of digital codes. The perceived bandwidth gain is much higher than just a single one or zero at a time.

      That way, he can access the mainframe any time, from any physical location, and it will translate directly into a virtual presence.

      • by roman_mir ( 125474 ) on Monday April 22, 2013 @06:53PM (#43520797) Homepage Journal

        They don't grt it. He solved the speed of processing and the lack of long term durability of storage by doing what's described in the original comment... Worked like a charm without needing to rithink the entire problem of a single bus used to retrieve and store data on the physical storage that still accessess data serially.

      • by crutchy ( 1949900 ) on Monday April 22, 2013 @07:48PM (#43521115)

        By spooling tractor feed paper between two printers in a loop, and by stopping and starting simultaneous paper-feed jobs, he can create a cybernetic feedback between the two printers that results in a series of quickly occurring "error - paper jam" messages that (due to two taped-down "reset" buttons) are quickly translated from the wide bandwidth analog physical matrix into kajamabits of digital codes

        i would be really careful doing that... the system may become self-aware

  • sounds like... (Score:2, Redundant)

    by stenvar ( 2789879 ) on Monday April 22, 2013 @06:48PM (#43520759)

    It sounds like he's doing standard GPU computations, loading everything into memory, and then calling it a "database", even though it really isn't a "database" in any traditional sense.

    • Re:sounds like... (Score:5, Informative)

      by tmostak ( 2904307 ) on Monday April 22, 2013 @07:04PM (#43520879)
      Hi, MapD creator here - and I have to disagree with you. The database ultimately stores everything on disk, but it caches what it can in GPU memory and performs all the computation there. So all the SQL operations are occurring on the GPU, after which, in case of the tweetmap demo, the results are rendered to a texture before being sent out as a png. But it works equally well as a traditional database - it doesn't do the whole SQL standard yet but can handle aggregations, joins, etc just like a normal database, just much faster. Todd
      • Re:sounds like... (Score:4, Interesting)

        by nebosuke ( 1012041 ) on Monday April 22, 2013 @07:33PM (#43521049)
        Just out of curiosity, did you use PGStrom [postgresql.org] or roll your own pgsql/GPU solution? If the latter, did you also hook into pgsql via the FDW interface or some other way?
      • by Anonymous Coward on Monday April 22, 2013 @07:41PM (#43521077)

        I'd be very interested to hear more details about the GPU SQL algorithms (JOIN in particular) if you are willing to share them. Did you use the set operations
        in Thrust or did you write something custom?

        Some of my colleagues are planning on releasing an open source library and some online tutorials about hash join and sort merge join in CUDA, and I would be very interested to share notes.

        • Re:sounds like... (Score:5, Informative)

          by tmostak ( 2904307 ) on Monday April 22, 2013 @07:58PM (#43521179)
          I'm not using thrust - I rolled my own hash join algorithm. This is something I still haven't optimized a great deal and I'm sure your stuff runs much better. Would love to talk. Just contact me on Twitter (@toddmostak) and I'll give you my contact details. Todd
      • by korgitser ( 1809018 ) on Monday April 22, 2013 @08:48PM (#43521447)

        I wonder what would it mean to the data if you were to lossily compress that png...

      • by stenvar ( 2789879 ) on Tuesday April 23, 2013 @02:35AM (#43522751)

        So, it sounds like you're implementing SQL as a data analytics language for in-memory data (plus a bunch of potentially useful algorithms), but apparently without the features that usually make a database a "database", like persistence, transactions, rollbacks, etc. It's those other features that make real databases slow, which is why you can't claim huge speedups over "databases" since you're not implementing the same thing.

        Data analytics on GPUs is a great thing, which is why tons of people are doing it. SQL isn't usually the language of choice because it isn't a good match and you have to build everything from scratch. GPU support in languages like R and Matlab gives you all the analytics features of SQL with a nicer syntax and really fast performance. Those languages also have tons of useful libraries for GIS, text analysis and visualization built in already.

  • by Anonymous Coward on Monday April 22, 2013 @06:51PM (#43520775)

    The 70x times seem optimistic. Does this include ALL the overheads for the GPU?
    But this done and patented over 2 years ago.
    http://www.scribd.com/doc/44661593/PostgreSQL-OpenCL-Procedural-Language

    And there has been earlier work using SQLite on GPU's.

    • by tmostak ( 2904307 ) on Monday April 22, 2013 @07:14PM (#43520935)
      The 70X is actually highly conservative - and this was benched against an optimized parallelized main-memory (i.e. not off of disk) CPU version, not say MySQL. On things like rendering heatmaps, graph query operations, or clustering you can get 300-500X speedups. The database caches what it can in GPU memory (could be 128GB on one node if you have 16 GPUs) and only sends back a bitmap of the results to be joined with data sitting in CPU memory. But yeah, if the data's not cached, then it won't be this fast. That's true, a lot of work has been done on GPU database processing - this is a bit different I think b/c it runs on multiple GPUs and b/c it tries to cache what it can on the GPU. Todd (MapD creator)
  • by __aaltlg1547 ( 2541114 ) on Monday April 22, 2013 @07:03PM (#43520867)
    The Egyptian government...
  • by BulletMagnet ( 600525 ) on Monday April 22, 2013 @07:12PM (#43520919)

    Still waiting for one/two....to play games on....

  • by crutchy ( 1949900 ) on Monday April 22, 2013 @07:53PM (#43521143)

    does it blend?

  • by Ghjnut ( 1843450 ) on Monday April 22, 2013 @08:05PM (#43521217)
    Maybe we should make it a habit of giving the owner some warning before slashdotting them. I know that if I ever get any concept development project up and running, I'm pretty excited to show my friends and tend to make it accessible before it's optimized enough to handle that king of onslaught.
  • by not_quite_a_user ( 2904313 ) on Monday April 22, 2013 @08:07PM (#43521225)
    Ingres was renamed to Actian and have released an analytic/reporting database called "Vectorwise" which makes use of SIMD and many other innovations in data throughtput techniques(everything in the Intel optimisation manual plus a lot more) and it gets more than 70 times performance. Check out TPC-H results "This is not an advertisement"
  • by zbobet2012 ( 1025836 ) on Monday April 22, 2013 @08:07PM (#43521229)
    While cool and all 125million tweets with geo tagging is at most: 1250000000*142bytes = 165 GB. That is not what "big data" considers a large data set. Indeed most "big data" queries are IO limited. For around 16k USD you can fit that entire working set in memory. You are not really in the "big data" realm into you have datasets in the 10's of TB's compressed (100's of TB's uncompressed).
    For these kinds of datasets, and where more compute is necessary there is MARs [gpgpu.org].
    • by rtaylor ( 70602 ) on Monday April 22, 2013 @09:26PM (#43521603) Homepage

      Agreed. That easily fits into memory (3 times actually) on our main OLTP DB.

      If it can fit into ram for less than $50K, it's not big data.

    • This project's innovation is noting that that GPUs have enough RAM now that you can push medium sized data sets into them if you have enough available. With lots of cores and high memory bandwidth, in-memory data sets in a GPU can do multi-core operations faster than in-memory data sets in a standard CPU/memory combination.

      That's great for simple data operations that are easy to run in parallel and when the data set is small enough to fit in your available hardware. Break any of those assumptions, and you've got a whole different set of problems to solve than what this is good for. I suspect none of those three requirements hold in the usual case for what people want out of "big data".

      • by tmostak ( 2904307 ) on Tuesday April 23, 2013 @01:22AM (#43522555)

        Hi - MapD creator here. Agreed, GPUs aren't going to me of much use if you have petabytes of data and are I/O bound, but what I think unfortunately gets missed in the rush to indiscriminately throw everything into the "big data bucket" is that a lot of people do have medium-sized (say 5GB-500GB) datasets that they would like to query, visualize and analyze in an iterative, real-time fashion, something that existing solutions won't allow you to do (even big clusters often incur enough latency to make real-time analysis difficult).

        And then you have super-linear algorithms like graph processing, spatial joins, neural nets, clustering, rendering blurred heatmaps which do really well on the GPU, which the formerly memory bound speedup of 70X turns into 400-500X. Particularly since databases are expected to do more and more viz and machine learning, I don't think these are edge cases

        Finally, although GPU memory will always be more expensive (but faster) than CPU memory, MapD already can run on a 16-card 128GB GPU ram server, and I'm working on a multi-node distributed implementation where you could string many of these together. So having a terabyte of GPU RAM is not out of the question, which, given the column-store architecture of the db can be used more efficiently by caching only the necessary columns in memory. Of course it will cost more, but for some applications the performance benefits may be worth it.

        I just think people need to realize that different problems need different solutions, and just b/c a system is not built to handle a petabyte of data doesn't mean its not worthwhile.

    • by KingMotley ( 944240 ) on Tuesday April 23, 2013 @11:29AM (#43526149) Journal

      125 million tweets != 1250000000

      125,000,000 * 142 = 16.5GB and easily fits into my desktop's memory.

  • For data processing workloads, a frequent problem with GPU acceleration is that the working dataset size is too large to fit into the available GPU memory and the whole thing slows to a crawl on data ingest (physical disk seeks, random much of the time) or disk writes for persisting the results.

    For folks serious about getting good ROI on their GPU hardware in real world scenarios, I strongly recommend you take a look at the fusion IO PCIe flash cards, which now support writing to and reading from them directly from CUDA via DMA, with little to no CPU handling required. (See: http://developer.download.nvidia.com/GTC/PDF/GTC2012/PresentationPDF/S0619-GTC2012-Flash-Memory-Throttle.pdf).

    I can't talk about what we do with it, but lets just say the following hardware combination has lead to interesting results;
    i) 16x PCIe slot chassis: http://www.onestopsystems.com/expansion_platforms_3U.php
    ii) 8x Nvidia Kepler K20x's
    iii) 8x Fusion IO 2.4TB IoDrive 2 Duo's

    We have been able sustain over 4 million data operations a second, each one processing ~16 K of data in a recoverable, transactionally consistent manner, totaling up to around 50 Gigabytes of data processed per second. All in a 5U deployment drawing less than 4 kilowatts.

  • by Gothmolly ( 148874 ) on Monday April 22, 2013 @09:52PM (#43521729)

    Granted its not free or cheap, but IBM will ship you a prebuilt rack of 'stuff' that will load 5TB/hour and scan 128GB/sec. PGStrom came out in the last year. Custom hardware/ASIC/FPGA for this sort of thing is not new.

  • by idbeholda ( 2405958 ) on Monday April 22, 2013 @10:23PM (#43521883) Journal
    As a data analyst/software engineer, it makes me glad to see these kind of actual strides are being made to ensure that both data and software will eventually start being designed properly from their inception. To have a single cluster database with anything more than a few thousand entries is nothing short of incompetence, and I believe anyone who does this should be publicly shamed and flogged. When dealing with excessively large amounts of data, it quickly becomes a necessity to have a paralleled database design to ensure that searches aren't hampered by long query times. It genuinely makes me thrilled to see someone else use this kind of design other than me, so when I put out numbers on my end, maybe my results won't seem as fantastical or unbelievable. Even though I don't know you personally, keep up the good work, Todd.
    • by tmostak ( 2904307 ) on Monday April 22, 2013 @11:09PM (#43522095)
      Thanks for the kind words! Hopefully this is just the start of a fun project... Todd
    • by b4upoo ( 166390 ) on Tuesday April 23, 2013 @12:51PM (#43527267)

      Most normal mortals won't have any knowledge of why a large database is useful. Frankly the first thing that leaps to my humble mind is
      trying to harvest money from the stock markets. Obviously there are numerous companies applying all kinds of computing power to the stock market. I do wonder if more computing power helps at this point or whether there is some toggle point at which massive data crunching would yield much better results.
                                    For example I don't know if people buy or sell more stocks if they are happy or if they are in a bit of suspense or under pressure. So how could a system that tries to compute the mood of the qualified buyers help me at all? It is rather like the car salesman's observation that the nonsense about gimmicks and hyping up the vehicle didn't sell many cars. The man told me that he signed deals when the buyers were worn out from going to dealership after dealership trying to decide what to do. They simply get so tired that they want relief and at that point can be sold most any old thing. How does one apply data to such complex beings as humans?

      • by idbeholda ( 2405958 ) on Wednesday April 24, 2013 @12:45AM (#43533741) Journal
        Map the human genome with a parallel database. The only "downtime" would be sequencing, but query times to test for different factors in a particular splice would be relatively short. The downside to this would be the amount of space required to group, and tie together relevant data. Something like this [tot-ltd.org] would probably be a start, which I still haven't gotten around to releasing in its entirety yet, given that I don't have much free time nowadays.
  • by WaffleMonster ( 969671 ) on Tuesday April 23, 2013 @12:06AM (#43522277)

    A: Indexes that don't suck.

    Using GPUs and massivelly parallel blah blah blah is cool and all but most databases are not processor limited so why should we care?

  • by biodata ( 1981610 ) on Tuesday April 23, 2013 @02:23AM (#43522709)
    40 million rows is what we used to manage in Oracle tables in the late 80s. Jeez, did this guy have no clue how to build a database?
  • by KPU ( 118762 ) on Tuesday April 23, 2013 @04:18AM (#43523065) Homepage

    Student writes inefficient code, learns how to optimize it using known techniques, it becomes faster. Film at 11.

  • by shikaisi ( 1816846 ) on Tuesday April 23, 2013 @04:44AM (#43523141)
    In Soviet Russia, GPU database creates you. Oh wait, wrong GPU [wikipedia.org]

"Everything should be made as simple as possible, but not simpler." -- Albert Einstein

Working...