I think there is still a long way toward predicting Hadoop job
runtime. The ICDE10 paper listed a lot of limitations of their methods
and is a short paper (4 pgs). In terms of the Berkeley research, based
on what I learned from a presentation done by Archana (the first
author of ICDE09 paper), they just scratched the surface of the
problem, and that the methodology seems to require quite exhaustive
experimentation on the configuration space to train their models -
overall, I am not convinced that the approach would work as well as
predicting sql query performance (the results they presented in the
ICDE09 paper).
-Hong
On Mar 2, 2010, at 6:09 PM, Jeff Hammerbacher wrote:
Predicting the run time of a MapReduce/Pig/Hive job has been
addressed by
folks at the University of Washington (e.g.
http://www.cs.washington.edu/homes/kmorton/ICDE10.pdf) and Berkeley
(e.g
using the techniques from
http://www.cs.berkeley.edu/~archanag/publications/ICDE09.pdf).
On Mon, Mar 1, 2010 at 4:48 PM, Edward Capriolo
<[email protected]>wrote:
I am looking at this many different ways.
For example: shuffle sort might run faster if we have 12 disks not
8 per
node.
So shuffle sort involves data size/ disk speed network speed/ and
processor speed/ number of nodes.
Can we find formula to take these (and more factors ) into account?
Once we find it we should be able to plug in 12 or 8 and get a result
close to the shuffle sort time.
I think it would be rather cool to have a long drawn out formula.that
even made reference to some constants, like time to copy data to
distributed cache,
I am looking at source data size, map complety, map output size,
shuffle sort time, reduce complexity, number of nodes and try to
arrive at a formula that will say how long a job will take.
From there we can factor in something like all nodes have 10 g
ethernet and watch the entire thing fall apart :)
On 3/1/10, brien colwell <[email protected]> wrote:
Map reduce should be a constant factor improvement for the algorithm
complexity. I think you're asking for the overhead as a function of
input/cluster size? If your algorithm has some complexity O(f(n)),
and
you spread it over M nodes (constant), with some merge complexity
less
than f(n), the total time will still be O(f(n)).
I run a small job, measure the time, and then extrapolate based on
the
bigO.
On 3/1/2010 6:25 PM, Edward Capriolo wrote:
On Mon, Mar 1, 2010 at 4:13 PM, Darren Govoni<[email protected]>
wrote:
Theoretically. O(n)
All other variables being equal across all nodes
should...mmmmm.....reduce to n.
That part that really can't be measured is the cost of Hadoop's
bookkeeping chores as the data set grows since some things in
Hadoop
involve synchronous/serial behavior.
On Mon, 2010-03-01 at 12:27 -0500, Edward Capriolo wrote:
A previous post to core-user mentioned some formula to
determine job
time. I was wondering if anyone out there is trying to tackle
designing a formula that can calculate the job run time of a
map/reduce program. Obviously there are many variables here
including
but not limited to Disk Speed ,Network Speed, Processor Speed,
input
data, many constants , data-skew, map complexity, reduce
complexity, #
of nodes......
As an intellectual challenge has anyone starting trying to
write a
formula that can take into account all these factors and try to
actually predict a job time in minutes/hours?
Understood, BIG-0 notation is really not what I am looking for.
Given all variables are the same, a hadoop job on a finite set of
data
should run for a finite time. There are parts of the process that
run
linear and parts that run in parallel, but there must be a way to
express how long a job actually takes (although admittedly it is
very
involved to figure out)
