HPC-day1

HPC-day1

by MADHURIMA CHAKRABORTY -
Number of replies: 7

Hello Bruno,

I have a few questions on the course you taught us yesterday:

1. Can the momentum balance equation for a Lagrangian tracer stated below be an example to an Embarrassingly Parallel Problem, as ever term in the equation is to be addressed separately?

M(dv/dt)= -M f X v +Ta + To + Ti + Fr + Fp

2. In slide-24, when you said 'the serial part dominates asymptotically', did you mean limP→∞ S = 1/f = 1 + Tp/Ts ? Would you kindly elaborate mathematically it a little further, please?

3. In slide-28, how did a 42% processor usage yield a speedup =3.4, and a 72% usage derived a speedup = 5.8? How did you do this conversion?

And have you already assigned the work to us? I am not being able to reach it by clicking on the assignment button.

Hope to see you tomorrow.

Thanks and best regards,

Madhurima.

In reply to MADHURIMA CHAKRABORTY

Re: HPC-day1

by Bruno Mundim -
Hi Madhurima,

apologies for my late replay. Regarding your questions:

1. I am not sure. Could you please state better what the bold face letters mean? Vectors? Tensors? Are you applying this equation to what? Photons? Particles? How are collisions modelled?

2. Yes, I meant mathematically exactly that. As we increase the number of cores the simulation speed-up with respect to the simulation using one core only approaches asymptotically the number 1/f, the inverse of the serial fraction. So for f=5%, the theoretical maximum speedup is 1/0.05 = 20. There is nothing you can do about it unless you change the algorithm in the serial part of the code or reorder how things are done to try to hide the serial part as much as possible. Of course, you could increase the system size. Usually the serial part doesn't grow with system size and that fraction goes down.

3. You can think of the speed up in the load balance context as \( S = T_1 / T_p = T_1 / (T_1 / (q p)) = q p \), where q is the percentage determined experimentally when all the tasks were run in parallel on p = 8 cores. So, initially q = 42% then S = 42% * 8 = 3.36. When the usage was 72% we had S = 72% * 8 = 5.76.

Regarding the assignment, it was posted yesterday morning. It is the same as the one on the slides.
In reply to Bruno Mundim

Re: HPC-day1

by MADHURIMA CHAKRABORTY -
Hello Bruno,
Thank you for your reply.
1. The bold letters stand for vector representation. And the equation is a simple classical mechanical equation of momentum balance for icebergs floated in the ocean currents as Lagrangian particles.
2. Right, okay.
3. That's fantastic.

I have a few more technical queries:
1. I am not being able to download my work using the 'scp' command yet. Will it be counted if I only copy the script in a word file and submit it to the course along with the answers to the questions asked?
1. I have an appointment for Biometric tomorrow at 12:15 pm. So, most likely I will miss tomorrow's session. Can I still claim the credit points if I submit the assignment by the stipulated time?
In reply to MADHURIMA CHAKRABORTY

Re: HPC-day1

by Bruno Mundim -
Hi Madhurima,

1. Are you modelling collisions of these ice bergs? If there is no dependency of one equation describing the motion of one iceberg and another one describing the motion of another iceberg, then it is embarrassingly parallel. I doubt it will be your case since they might collide, for example. However I am not fully aware of all assumptions. The point is if there is data dependencies or not.

What is the problem you are facing with scp? You can copy the script to a .txt file if you are working on windows (using notepad for example) and submit it. That will be counted as work done.
And yes, you can claim credits if you submit your assignment. I will post the recordings afterwards.
In reply to Bruno Mundim

Re: HPC-day1

by MADHURIMA CHAKRABORTY -
Hello Bruno,

No, they do not collide as they either move jam-packed with the sea-ice in Winter, or separated by few kilometers in Summer. Nonetheless, I got the answer, thank you.

scp lcl_uothpc101s2101@teach.scinet.utoronto.ca:/scratch/l/lcl_uothpc101/lcl_uothpc101s2101/introhpc/out C:\Users\matmo\Downloads\Scinet_assignment

I created a folder in Downloads named as Scinet_assignment and was trying to download the work using the command above. I got the error saying:

get_socket_address: getnameinfo 8 failed: Name or service not known
get_socket_address: getnameinfo 8 failed: Name or service not known
get_socket_address: getnameinfo 8 failed: Name or service not known
get_socket_address: getnameinfo 8 failed: Name or service not known
get_socket_address: getnameinfo 8 failed: Name or service not known
get_socket_address: getnameinfo 8 failed: Name or service not known
get_socket_address: getnameinfo 8 failed: Name or service not known
get_socket_address: getnameinfo 8 failed: Name or service not known
Password:
ssh: Could not resolve hostname c: Name or service not known
lost connection
Connection to teach.scinet.utoronto.ca closed.

Thanks,
Madhurima.
In reply to MADHURIMA CHAKRABORTY

Re: HPC-day1

by Bruno Mundim -
On your compute navigate to the directory you want to save your work:

C:\Users\matmo\Downloads\Scinet_assignment

Now you can issue the following command to get all your directory on teach cluster with your work:
scp -r lcl_uothpc101s2101@teach.scinet.utoronto.ca:/scratch/l/lcl_uothpc101/lcl_uothpc101s2101/introhpc/'*' .

Note the dot at the end. It means current directory in linux. What software are you using in windows? WinScp might work for you, or the MobaXterm terminal.
In reply to Bruno Mundim

Re: HPC-day1

by MADHURIMA CHAKRABORTY -
Hello,

That's great. I signed out from the node. I shall try to download the work tomorrow, I am sure it'll work. I am using MobaXterm like you suggested in the first class.

Thank you so much for the course. This workshop was extremely engaging. It's a shame I'm going to miss the closure. Will the forum be still open after tomorrow's class, in case I have some more questions after I go through the recording?

Best,
Madhurima.
In reply to MADHURIMA CHAKRABORTY

Re: HPC-day1

by Bruno Mundim -
You are welcome. I am glad you found it engaging.

Regarding the forum, yes, I believe you can still post as long as you had registered for this course. However this site's configuration is sometimes not so simple to grasp. Therefore if you try to post and are not able to do so, please send us a message to courses@scinet.utoronto.ca instead. If your question is related to system issues or software installation or any other issue not related to this course material, please reach us support@scinet.utoronto.ca.

Regards,
Bruno.