Scaling Research Computing Workflows

You are not logged in.
Please Log In for full access to the web site.
Note that this link will take you to an external site (https://petrock.mit.edu) to authenticate, and then you will be redirected back to this page.

Review the concepts you learned from Scaling Research Computing Workflows module

1) Analyzing the Serial Application

You have a workflow that processes a large number of files, spread over many sub-directories. Each file is processed independently but requires 2x more than the nominal RAM per core on the hardware. What canonical workflow does this processing map to?

High Throughput

Loosely Coupled

Parallel

You are using a Python package to train a machine learning model. What do you need to check during your serial testing? (Check all that apply)

confirm that the package is installed and loaded properly

determine whether or not the package uses threads

check to see whether or not accelerators are supported

time the code to get a serial baseline for time

determine how much memory is required in order to request the appropriate resources

Which of the following techniques are recommended to uncover the computational bottlenecks? (Select all that apply.)

computing the size of the data arrays

timing the entire code

calculating the number and size of the input files

timing major functions

using a commercial profiler for the code, when available

researching the hardware that the code will run on

What should you focus on when determining your resource requirements? (Select all that apply.)

the time it takes to run to solution

the amount of memory required to complete the code

the hardware components required by the code, CPU or GPU

the size of the input files

the size of the output files

the level of independence and constraints on sharing memory

2) Performance

What is the limiting factor for speed-up?

Number of processors

Serial portion of the code

Network speed/communication latency

Size of the messages

Parallel portion of the code

What 2 actions need to be taken to achieve maximum speed-up for any application? (Select all that apply.)

Minimize the communication overhead

Adjust the network communication parameters

Scale back the memory footprint

Optimize the serial portion

Will more processors always lead to faster time to solution?

Yes

3) Tools for Application Analysis

Which command can be used to show statistics about your jobs, in particular to get the peak memory usage? Enter just the name of the command, without any options, flags, or arguments.:

Which command can be used to show information about the current CPU and memory utlization, processes and threads, and load? Enter just the name of the command, without any options, flags, or arguments.

Which command would you use to show statistics on current GPU utilization? Enter just the name of the command, without any options, flags, or arguments.

True or false: The MaxRSS and MaxVMSize fields to used show memory usage in sacct are only accurate for completed jobs.

True

False

What must you do before running the nvidia-smi or htop commands to evaluate one of your jobs?

Nothing

ssh to the node where the job is running.

Make sure your job has completed.

ssh to the login node where you lauched the job.