1. Providing a Supported Online Course on Parallel Computing Steven I
Providing a Supported Online Course on Parallel Computing Steven I. Gordon, James Demmel, Susan Mehringer, Jay Alameda, Razvan Carbunescu

2. Class Overview
Interest in providing a full semester course to a large audience
Opportunity for collaborative effort to provide the course online
Faculty and students at University of California Berkeley
XSEDE education program
Training staff from Cornell
ESTEO support staff
Created full online course open to the world and advertised through the XSEDE training portal

3. Course Lectures

4. Conversion for Online Use
Editing videos into smaller segments
Added reference links and glossary
Adding quizzes to test comprehension
Devising a grading scheme
Policy on multiple attempts
Providing mechanism for support
Moodle server with forum monitored by course TA and XSEDE ESTEO staff

5. Lecture and Moodle Sites

6. Logistical Issues – Part 1
Finding appropriate XSEDE platforms for programming exercises
Possibility of using different platforms for assignments to take advantage of diversity of computer architectures
Testing of codes
Creation of autograders to handle potentially large number of students

7. Logistical Issues – Part 2
Large number of people initially interested
Original limitation set at 300
Opened a waiting list
Accepted 345 students
Setting up accounts
Connection to Moodle server using XSEDE portal credentials
Moodle server used to turn in homework and ask for help
Fewer than 100 actually took this step
Accounts on XSEDE machines require one by one entry
Who would actually continue to participate and when to set up accounts became an issue
Panic by XSEDE service providers

8. Map of All Participants

9. Resulting Course
Videos and quizzes in the Cornell Virtual Workshop and available to anyone with an XUP account
376 people used the first lecture
145 of the 345 students active on that site
Major drop-off in participation over time
Programming homework
Major drop in participation for homework assignments
36 – assignment 1 – Optimizing Matrix Multiply – used 200 CPU hours
23 – assignment 2 – Toy Particle Simulation – used CPU hours
18 – assignment 3 – Knapsack (on PGAS language) – used CPU hours
CPU hours presented are those charged (actual time was likely underestimated on HW1 and overestimated for HW2)

11. Grading and Support Issues
Autograder generally split into correctness and performance
Both code parts provided to students through initial files & scripts but unmodified files used for grading
Most submissions received within 48 hours of deadline creating high loads for short times
Very high latency (30+ minutes) for autograder to give result for runs of HW2&3 due to scaling tests
A better integration/automization for submitted files is needed to remove compilation, naming issues etc.
Small changes done to autograders at students requests (Ex: using different OpenMP thread allocation policies - compact)

12. Certificate of Completion
Only 18 completed all of the materials with a sufficiently high performance to earn the final certificate
All others dropped out at one point or another
To receive certificate a combined grade above the total given by a passing grade on each quiz and assignment was required

13. Evaluation – Students with Certificate
Population (16 of 18 replies)
75% graduate students; 13% faculty; 6% post-docs; 6% other
% Responses to post course survey (Strongly agree, agree, neutral, disagree, strongly disagree)
Question SA A N D SD
Goals for taking this course achieved 25 56 19
Course well organized 31 63 6
Course content thoroughly covered 50 13
Quizzes of appropriate difficulty 44
Homework of appropriate difficulty
Background sufficient
Happy for time invested
Satisfied with online delivery

14. Evaluation - Others Population (56 of 325 replies)
68% Grad students; 9% undergrad; 7% faculty; 8% staff and other
Most enrolled for: extend knowledge of the topic (75%), generally interested in the topic (43), professional development (23), interested in how online courses taught (18), to supplement another class (13)
Not completed because: Amount of time required (88%), insufficient background knowledge (30), technological difficulties (16), never intended to complete (9)
Take another course in this format: Yes (91%) No (4%)
Interested in other XSEDE opportunities: Yes (82%) No (5%)
Plan on applying for XSEDE allocation: Yes (30%) No (18%) Have one (18%)

15. Issues for Future Offerings
Possible ways to sustain participation
Change the timing of the class to minimize conflicts with regular course activities
Formalize participation by providing credit at student’s home institutions
Staggered start dates to minimize number of students on XSEDE resources at one time
Divide course into two levels of completion
Improved focus on serious students
Make first assignment offline through use of virtual machine
Create XSEDE accounts for only those successfully completing first assignment
Spread programming assignments across multiple XSEDE machines
Credit course with local supervising instructor

16. Acknowledgements
National Science Foundation NSF OCI
XSEDE TEOS Evaluators – Lizanne Destefano and Lorna Rivera
XSEDE Resource Providers – Pittsburgh Supercomputing Center, Texas Advanced Computing Center