Re: Large Data Sets in Julia

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view
|

Re: Large Data Sets in Julia

André Lage
hi Viral,

Do you have any news on this?

André Lage.

On Wednesday, July 3, 2013 at 5:12:06 AM UTC-3, Viral Shah wrote:
Hi all,

I am cross-posting my reply to julia-stats and julia-users as there was a separate post on large logistic regressions on julia-users too.

Just as these questions came up, Tanmay and I have been chatting about a general framework for working on problems that are too large to fit in memory, or need parallelism for performance. The idea is simple and based on providing a convenient and generic way to break up a problem into subproblems, each of which can then be scheduled to run anywhere. To start with, we will implement a map and mapreduce using this, and we hope that it should be able to handle large files sequentially, distributed data in-memory, and distributed filesystems within the same framework. Of course, this all sounds too good to be true. We are trying out a simple implementation, and if early results are promising, we can have a detailed discussion on API design and implementation.

Doug, I would love to see if we can use some of this work to parallelize GLM at a higher level than using remotecall and fetch.

-viral

On Tuesday, July 2, 2013 11:10:35 PM UTC+5:30, Douglas Bates wrote:
On Tuesday, July 2, 2013 6:26:33 AM UTC-5, Raj DG wrote:
Hi all,

I am a regular user of R and also use it for handling very large data sets (~ 50 GB). We have enough RAM to fit all that data into memory for processing, so don't really need to do anything additional to chunk, etc.

I wanted to get an idea of whether anyone has, in practice, performed analysis on large data sets using Julia. Use cases range from performing Cox Regression on ~ 40 million rows and over 10 independent variables to simple statistical analysis using T-Tests, etc. Also, how does the timings for operations like logistic regressions compare to Julia ? Are there any libraries/packages that can perform Cox, Poisson (Negative Binomial), and other regression types ?

The benchmarks for Julia look promising, but in today's age of the "big data", it seems that the capability of handling large data is a pre-requisite to the future success of any new platform or language. Looking forward to your feedback,

I think the potential for working with large data sets in Julia is better than that in R.  Among other things Julia allows for memory-mapped files and for distributed arrays, both of which have great potential.

I have been working with some Biostatisticians on a prototype package for working with snp data of the sort generated in genome-wide association studies.  Current data sizes can be information on tens of thousands of individuals (rows) for over a million snp positions (columns).  The nature of the data is such that each position provides one of four potential values, including a missing value.  A compact storage format using 2 bits per position is widely used for such data.  We are able to read and process such a large array in a few seconds using memory-mapped files in Julia.  The amazing thing is that the code is pure Julia.  When I write in R I am always conscious of the bottlenecks and the need to write C or C++ code for those places.  I haven't encountered cases where I need to write new code in a compiled language to speed up a Julia function.  I have interfaced to existing numerical libraries but not writing fresh code.

As John mentioned I have written the GLM package allowing for hooks to use distributed arrays.  As yet I haven't had a large enough problem to warrant fleshing out those hooks but I could be persuaded.

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.