DOC PREVIEW
Berkeley COMPSCI C267 - CS 267 Homework Assignment

This preview shows page 1 out of 3 pages.

Save
View full document
View full document
Premium Document
Do you want full access? Go Premium and unlock all 3 pages.
Access to all documents
Download any document
Ad free experience
Premium Document
Do you want full access? Go Premium and unlock all 3 pages.
Access to all documents
Download any document
Ad free experience

Unformatted text preview:

Assignment 0 Biography My name is Razvan Corneliu Carbunescu and I am a first year graduate student within the CS Department here at UC Berkeley. I’m Romanian but I did my undergraduate degree in the US also at Louisiana State University. My research interests would fall under the scope of Scientific Computing although I’m generally interested in all of the aspects present within ParLab: Architecture design, Machine learning, etc. (Perhaps the least in verification). Although I have programmed code before that was run on a supercomputer I did so within a framework from LSU named Cactus. Therefore one of the things I hope to get out of this class is the ability to create and run a program efficiently on a parallel computer relying solely on basic communication protocols like MPI. Parallel Application – Introduction The application which I decided to focus on is the analysis of very large Molecular Dynamics Simulation Trajectories. (up to Terascale size datasets). This work was presented at SC’08 and addresses the need for a reinvention of serial analysis of MD simulations. This problem is a simple subset of the need in general for analysis, visualization or input/output tools to also become parallel as the numerical simulations driving them have reached the point of delivering/requiring huge data sets. Below are two MD simulations and two examples of analysis that can be done on such sets a) Ion permeation through a channel b) Analyzing from MD simulations electron density maps for an area. Parallel Application – Method The method used to parallelize the analysis of the frames from MD simulations is based on the idea of decoupling the analysis of each individual frame from the analysis of coupling between them. Since after this decoupling the frames become independent for the first step of theanalysis process the framework can use parallelism to split up work evenly amongst processors (cores) and then combine the data either at the end or at a step where intermediate results are necessary. The way this is done is by an extension of the MapReduce concept based on a generalization of the functions that map and reduce should provide. Also another generalization is made to allow for multiple levels of MapReduce to be applied to the analysis. The way data is communicated between processors is by the use of MPI and communicating in groups of two rather than maintaining a single master node (which is the method used in MapReduce). Parallel Application – Target Platform It is clear from the paper presented that the application is targeted for distributed memory going so far as abstracting that cores on the same chip are individual processors. The Application is also targeted towards distributed memory because of the concept of ‘quickly’ running out of memory on a chip for Terascale simulations so the framework is built to be able to handle input from a separate storage location. Parallel Application – Results The framework presented got good results in the scaling studies performed on a local LINUX cluster in both sample problems used. The first problem is the permeability problem mentioned above and the second is a more complex example requiring two levels of MapReduce for the framework. As can be seen the code obtained significant improvement to 2 orders of magnitude over the serial code on up to 512 cores for both examples. For the second sample problem weak scalingresults are also available with the code handling the analysis of 1 Terabyte of data within 15 minutes. Parallel Application – Problems There are a couple of problems that I see with the application framework presented here. The biggest problem by far is the reliability of the code since because of the communication algorithm presented “if one node goes down during an analysis execution, the HiMach user program must be re-executed from the beginning”[1] (HiMack is the name of the framework). This is especially worrisome since if scaling is expected to continue tolerance should be built into the model to account for failures in cores. Another concern is the fact that for most analysis it seems like the actual gathering of the interpreted data from all frames (reduce step) will still result in a serial bottleneck regardless of the processing speed of the individual frames. Another aspect that I considered most surprising is their statement that the domain specialist needs to write only “serial code” within this framework and that the framework will take care of all the parallelism. While it is true that the programmer does not need to worry about setting up the MPI calls or all the other internals I think that the design of correct map/reduce functions which keep track and resolve data dependencies across processors is non-trivial and definitely not absent. Parallel Application – Conclusions While I believe that the implementation of the framework still has many flaws it is a great improvement over the previous state of tools available and is a great parallel application in process of development. Parallel Application – References [1] A scalable parallel framework for analyzing terascale molecular dynamics simulation trajectories Tiankai Tu, Charles A. Rendleman, David W. Borhani, Ron O. Dror, Justin Gullingsrud, Morten Ø. Jensen, John L. Klepeis, Paul Maragakis, Patrick Miller, Kate A. Stafford, David E. Shaw Article No. 56, SC’08: Proceedings of the 2008 ACM/IEEE conference on


View Full Document

Berkeley COMPSCI C267 - CS 267 Homework Assignment

Documents in this Course
Lecture 4

Lecture 4

52 pages

Split-C

Split-C

5 pages

Lecture 5

Lecture 5

40 pages

Load more
Download CS 267 Homework Assignment
Our administrator received your request to download this document. We will send you the file to your email shortly.
Loading Unlocking...
Login

Join to view CS 267 Homework Assignment and access 3M+ class-specific study document.

or
We will never post anything without your permission.
Don't have an account?
Sign Up

Join to view CS 267 Homework Assignment 2 2 and access 3M+ class-specific study document.

or

By creating an account you agree to our Privacy Policy and Terms Of Use

Already a member?