In my tests, ABt-job took under 3 minutes per mapper and practically
no time for reducing. so it should be running at about 4 minutes on a
cluster with sufficient capacity (in your case, something like 10=11
nodes, it seemed). Ok i'll rerun in our QA on Monday again to see
what's happening.

On Sat, Dec 17, 2011 at 3:04 PM, Dmitriy Lyubimov <[email protected]> wrote:
> ABt-job  37mins, 22sec
> this guy should run under Bt-job (under 9 minutes in your case) i
> think . In my tests it was. is this with 922 patch?
>
> And it should be mentioned that the cluster size couldn't accomodate
> all the generated tasks, is this correct assessment?
>
>
> On Sat, Dec 17, 2011 at 2:58 PM, Sebastian Schelter <[email protected]> wrote:
>> On 17.12.2011 17:27, Dmitriy Lyubimov wrote:
>>> Interesting.
>>>
>>> Well so how did your decomposing go?
>>
>> I tested the decomposition of the wikipedia pagelink graph (130M edges,
>> 5.6M vertices making approx. quarter of a billion non-zeros in the
>> symmetric adjacency matrix) on a 6 machine hadoop cluster.
>>
>> Got these running times for k = 10, p = 5 and one power-iteration:
>>
>> Q-job    1mins, 41sec
>> Bt-job   9mins, 30sec
>> ABt-job  37mins, 22sec
>> Bt-job   9mins, 41sec
>> U-job    30sec
>>
>> I think I'd need a couple more machines to handle the twitter graph
>> though...
>>
>> --sebastian
>>
>>
>>> On Dec 17, 2011 6:00 AM, "Sebastian Schelter" <[email protected]>
>>> wrote:
>>>
>>>> Hi there,
>>>>
>>>> I played with Mahout to decompose the adjacency matrices of large graphs
>>>> lately. I stumbled on a paper of Christos Faloutsos that describes a
>>>> variation of the Lanczos algorithm they use for this on top of Hadoop.
>>>> They even explicitly mention Mahout:
>>>>
>>>> "Very recently(March 2010), the Mahout project [2] provides
>>>> SVD on top of HADOOP. Due to insufficient documentation, we were not
>>>> able to find the input format and run a head-to-head comparison. But,
>>>> reading the source code, we discovered that Mahout suffers from two
>>>> major issues: (a) it assumes that the vector (b, with n=O(billion)
>>>> entries) fits in the memory of a single machine, and (b) it implements
>>>> the full re-orthogonalization which is inefficient."
>>>>
>>>> http://www.cs.cmu.edu/~ukang/papers/HeigenPAKDD2011.pdf
>>>>
>>>> --sebastian
>>>>
>>>
>>

Reply via email to