5 ms·
Official Google Blog: Sorting 1PB with MapReduce
- jwilliams 18y agoIf my math is right: (68sec * 1024) / 60 / 60 / 4 --- (68 sec for 1TB) (1024 to make 1TB) /60/60 (hours) and /4 as there are 4x as many computers.... So ideally taking the 1TB sort to 1PB on 4x the hardware would be 4.83hrs. Google's 1PB sort was in 6 hours - So that's fairly linear. Impressive.
- jmtulloss 18y agoThe article implied they were using hard disk sizes. IE. 1 Petabyte is 1000 Terabytes. Not that it makes that much of a difference.
- lutorm 18y agoIsn't the mapreduce sort NlogN in the number of entries? That would make 1e13 log 1e13/(1e10 log 1e10) 68s/4 = 6.1h. So they were actually faster than the possible scaling? Doesn't sound right...
- russell 18y agoSounds dead on to me. The typical sort is O(N logN). O can be slightly different between the 1T runs and 1P runs, e.g. slightly difference latency or pattern of disk failures. Every sort is sensitive to the data distribution. There is probably an interesting distribution of results across the runs.
- bd 18y agoThey mention that they did some extra optimizations ("stragglers") for petabyte test (that were not there yet for the terabyte test).
- jwilliams 18y agoThis is a really good point. I was assuming that the sort was actually a partial sort - returning the top "x" (admittedly - even then isn't linear). However, I was also assuming (n << x), which would almost certainly be the case with a large data set. It's a fairly common approach, but not sure why I automatically assumed it. If the sort is complete, then yeah, that changes things.
- est 18y agoOK. Zhuangbility.
- iowahansen 18y ago4000 computers over 6 hours... Makes you wonder about electricity costs of this experiment.
- wmf 18y agoSpeaking of electricity, at one point Google claimed that during a Web search your PC uses more energy sending the request, waiting, and displaying the results than Google uses to perform the search. So the next time you need to sort 1PB of data it may be cheaper to send it to Google. :-)
- gsmaverick 18y agoLol, good idea. Although I could see M$ or Yahoo! looking to make some sort of a deal like that.
- slackerIII 18y agoHeh... I guess this is what Yahoo gets for talking smack about how quickly they could sort 1TB: http://developer.yahoo.net/blogs/hadoop/2008/07/apache_hadoop_wins_terabyte_sort_benchmark.html http://developer.yahoo.net/blogs/hadoop/2008/07/apache_hadoo...
- sh1mmer 18y agoGoogle have been working on Map Reduce for 4 years longer than anyone else. I'm actually pretty happy that we (Yahoo!) are talking about what we are doing in the Open and contributing in an Open project. You could reproduce what we did. Obviously you can't reproduce what Google did. What's more useful?
- nebula 18y agoWe were writing it to 48,000 hard drives (we did not use the full capacity of these disks, though), and every time we ran our sort, at least one of our disks managed to break (this is not surprising at all given the duration of the test, the number of disks involved, and the expected lifetime of hard disks). I'm a bit at lost here. Does that mean that one in ~50K hard disks fail within six hours of usage?
- russell 18y agoNot at all surprising given the large number of disks involved. Here is a very interesting analysis of actual Google failure rates: http://storagemojo.com/2007/02/19/googles-disk-failure-experience/ http://storagemojo.com/2007/02/19/googles-disk-failure-exper...
- nebula 18y agoThanks for the pointer; it's interesting.
- wmf 18y ago1 failure per 50K in 6 hours is the same as 1 per 34 in a year, or a 3% annual failure rate. I think this is the same number Google reported in their disk failure paper.
- nebula 18y agoActually, I did this math before I posted my comment. What I don't understand is this: With x% annual failure rate, are the failures uniformly distributed over time? I would expect that it's probably a normal distribution over time. (i.e., no failures to very few failures in the beginning, and the failures keep increasing with time).
- bd 18y agoIt's probable that they didn't use 48,000 brand new hard drives, but reused whatever they already had in their datacenters. Which would have a mixture of different drive ages, I guess.