Official Google Blog: Sorting 1PB with MapReduce

http://googleblog.blogspot.com/2008/11/sorting-1pb-with-mapreduce.html

by Anon84 • 18 years ago

57 19 18 years ago

googleblog.blogspot.com

jwilliams2• 18 years ago

If 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• 18 years ago

The article implied they were using hard disk sizes. IE. 1 Petabyte is 1000 Terabytes. Not that it makes that much of a difference.

lutorm• 18 years ago

Isn'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...

russell2• 18 years ago

Sounds 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.

bd2• 18 years ago

They mention that they did some extra optimizations ("stragglers") for petabyte test (that were not there yet for the terabyte test).

jwilliams2• 18 years ago

This 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• 18 years ago

OK. Zhuangbility.

iowahansen• 18 years ago

4000 computers over 6 hours... Makes you wonder about electricity costs of this experiment.

wmf2• 18 years ago

Speaking 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• 18 years ago

Lol, good idea. Although I could see M$ or Yahoo! looking to make some sort of a deal like that.

slackerIII• 18 years ago

Heh... 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_hadoo...

sh1mmer• 18 years ago

Google 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?

nebula4• 18 years ago

We 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?

russell2• 18 years ago

Not 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-exper...

nebula4• 18 years ago

Thanks for the pointer; it's interesting.

wmf2• 18 years ago

1 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.

nebula4• 18 years ago

Actually, 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).

Official Google Blog: Sorting 1PB with MapReduce

Related Stories