Slashdot Mirror
Google Sorts 1 Petabyte In 6 Hours
krewemaynard writes "Google has announced that they were able to sort one petabyte of data in 6 hours and 2 minutes across 4,000 computers. According to the Google Blog, '... to put this amount in perspective, it is 12 times the amount of archived web data in the US Library of Congress as of May 2008. In comparison, consider that the aggregate size of data processed by all instances of MapReduce at Google was on average 20PB per day in January 2008.' The technology making this possible is MapReduce 'a programming model and an associated implementation for processing and generating large data sets.' We discussed it a few months ago. Google has also posted a video from their Technology RoundTable discussing MapReduce."
for knowing how important the Library of Congress metric is to us nerds!
Yay! We finally have unit conversion from 1 LoC to bytes! So...20 PB = 6LoC, means that 1 LoC = 3,333... PB :)
Consider a data set of two numbers, each .5 petabyte big. It should only take a few minutes to sort them and there's even a 50% chance the data is already sorted.
Sorts have been parallelized and distributed for decades. It would be interesting to benchmark Google's approach against SyncSort. SyncSort is parallel and distributed, and has been heavily optimized for exactly such jobs. Using map/reduce will work, but there are better approaches to sorting.
I will be able to catalog my pr0n in my lifetime:
Blondes, Brunettes, Red heads, Beastial^H^H^H^H^H "Other"
No sig for you!!
I realize, slashdot..., but maybe you could glance at the article which states:
10 trillion 100-byte records
It's not enough to sort by blond, black, gay, scat, etc. Some categories are a combination that don't belong in a hierarchy.
That is where tagging comes in. Sorting can be done on-the-fly, with no one category intrinsically more important.
I suggest you read Slashdot
Not a big deal, that's just the data they have on you.
lol: You see no door there!
Sorting a petabyte sounds pretty impressive, but I don't think it was a whole yotta work.
Bad puns gave me bad karma. =(
Today from Google, the god of all things and doer of all things good in the universe, many millions of dollars in computer equipment were able to sort lots of things, in about the amount of time you would think it would take for millions of dollars of equipment to sort things.
In other news, a woodchuck was found chucking wood as fast as a woodchuck could chuck wood.
Congrats Google, you have a HUGE data set, and an even bigger wallet.
- Adam L. Beberg - The Cosm Project - http://www.mithral.com/
Exactly. It's unclear if their better time was a software engineering or algorithmic feat, though. Hadoop was able to finish sorting the 1 TB benchmark dataset in 209 s; TFA states Google pulled the same event off in 68 s. The Yahoo blog post you linked to says their compute nodes each sported 4 SATA HDDs. Note TFA mentions Google's 1 PB dataset sort used 48,000 HDDs split between 4,000 machines, or 12 HDDs to a machine. If Google used the same machines to perform their 1 TB sort, then they had 3 times as many HDDs on each compute node, and could probably pull data from storage 3 times as fast. 209 s / 68 s ~ 3.1 -- coincidence, or not? =)
The individual functions map and reduce are quite standard. The innovation here is the systems work they've done to make it work on such a large scale. All the programmer needs to worry about is implementing the two functions, they don't have to worry about distributing the work, ensuring fault tolerance, or anything else for that matter. That is the innovation.
They mention in the article that if you try and sort a petabyte you WILL get hard disk and computer failures. Hell, you can only read a terabyte hard disk a few times before you encounter unrecoverable errors. The system for executing those maps and reduces is what is important here. The important parts are in the design details, such as dealing with stragglers. If you have 4000 identical machines, you won't necessarily get equal performance. If a few of those machines have a bit flipped and started without disk cache, they might see a huge decrease in read/write performance. The system needs to recognize this and schedule the work differently. That can make a huge difference in execution time. If you graph the percentile complete of a MR job, you'll often see that it quickly reaches 95% and then plateaus. The last 5% may take 20% of the time, and good scheduling is required to bring this time down.
But like I said, the innovation isn't in the idea of using a Map and Reduce function, it is the system that executes the work.