CS PhD Profile Evaluation

[anacron]

Hey guys,

I'm going to be applying for PhD programs during Fall 2011 and I just wanted to get your opinions on where my profile stands.

A little background, I am double majoring in Computer Science and Computer Engineering. My research interests are in Cloud and High-Perf. Computing.

Here are some key statistics:

Cum. GPA : 3.86/4.00

Major GPA's : have not calculated the exact numbers but definitely at or above 3.9.

Research: Currently working on the performance evaluation of high performance computing applications in virtualized environments. Will be submitting a project proposal to use the Magellan cloud testbed at NERSC Lawrence Berkeley lab. This will go towards my senior thesis as well as a journal paper around January.

Recently admitted to UIUC's summer REU and will spend this summer working on a research project culminating in a poster presentation.

Publications: One paper submitted to a journal, should receive decision soon(Invited, peer-reviewed journal). I am currently working on another paper which would be a first author paper.

LoRs: 2 from my research advisors, one of whom is quite well known in High-Perf Computing. I have worked with both for all of my research projects and have done well in classes they were teaching. Hopefully 1 more from the PI at the UIUC REU and the 4th(for fellowships) from my school's undergraduate director.

Awards: Dean's list, scholarships, academic related awards. etc.

Others:

Teaching assistant for an intro. CS class.

Teaching assistant for an electrical engineering class at John's Hopkins summer program

Math Tutoring

It would be great if you guys could give me some advise on where I could apply and if I should skew my list toward the Top 20 or keep more safety schools.

Edited April 24, 2011 by anacron

[frenzydude]

I think you should definitely target the top 20. Also, I think you meant Fall 2012 and not 2011.

[anacron]

I think you meant Fall 2012 and not 2011.

I will be applying for admission into the Fall 2012 cycle. The actual applications will be submitted this November/December (i.e. Fall 2011).

Thanks for replying!

I will be taking my GRE's this summer and am not too worried about the scores. I feel that I have shown my capabilities with my research and publications. Of course, I don't plan on bombing on the tests either

[frenzydude]

I will be applying for admission into the Fall 2012 cycle. The actual applications will be submitted this November/December (i.e. Fall 2011).

Thanks for replying!

I will be taking my GRE's this summer and am not too worried about the scores. I feel that I have shown my capabilities with my research and publications. Of course, I don't plan on bombing on the tests either

Yup.. doesn't matter much. Instead try to do some good work in the summer and get recos/papers.

[OH YEAH]

It's really hard to evaluate your chances when all of you provide are basic statistics!

Journals are not the primary means of publication in CS -- which journals did you get accepted to, and how prestigious are they? Is there any reason you guys didn't submit to a conference, especially since cloud computing is such a hot and quickly moving field?

Your research: were these your ideas? Was your job to swab the decks, or did you do stuff worth talking about? I did two REUs, and while I didn't get published because of them (for various, unfortunate reasons ), I went into it with the expectation of publishing a paper. Why are you aiming for just a poster presentation (these don't count for much?)

It's good that your LoR writers are well known, but that doesn't really say anything in and of itself. How well do you know them? How long have you known them? What do they think of you, and how would they compare you to others in your cohort?

Which schools are you interested in (based on the professors, not their USNews ranking, although that is important too)?

[anacron]

It's really hard to evaluate your chances when all of you provide are basic statistics!

Journals are not the primary means of publication in CS -- which journals did you get accepted to, and how prestigious are they? Is there any reason you guys didn't submit to a conference, especially since cloud computing is such a hot and quickly moving field?

Your research: were these your ideas? Was your job to swab the decks, or did you do stuff worth talking about? I did two REUs, and while I didn't get published because of them (for various, unfortunate reasons ), I went into it with the expectation of publishing a paper. Why are you aiming for just a poster presentation (these don't count for much?)

Although it is true that I didnt specify which journal or its prestige would this really make the difference with addcomms? But, if it maybe of use, we submitted to the Journal of Computer Systems Science and the first author paper is targeted at IEEE Computer.

The research proposal was my idea and will lead towards another first author paper if we make substantial progress as well as contribute towards my senior thesis, its scope is comparable to an MS thesis and I will be working towards that target. My advisors were very excited about the proposal and I will be working on it starting the fall term.

I have worked closely with both of my advisors, taken classes with them and have regular contact with them. They are the main reason why I have gained a lot of good opportunities for this summer. Out of which I am choosing to do the UIUC REU. The poster presentation was a default at the REU, The PI and I have not discussed projects yet. (they are still sending out acceptances) I would like to present the work at a conference etc. but that would depend on the scope/novelty of the project given the 10-week deadline.

Regarding top schools:

UIUC is a major one since they made good contributions in Cloud computing including Open Cirrus.

Stanford for their work with the Magellan testbed

Berkeley, RAD Labs

U Washington

U Texas Austin

Several safety schools

On a side note, the reason for journal publications are because my advisor needs more journals. I have no problems with conferences if my advisors are ok. After all I am an undergrad and will have 5-6yrs of PhD to publish to conferences

My advisors have not pushed me to "swab the decks" etc and have let me have complete freedom with my projects. (except for the JCSS paper which was my first contribution and I did not have the background knowledge at the time).

Hopefully this gives a better idea? I would love to hear your thoughts.

Edited April 24, 2011 by anacron

[adinutzyc]

You have a VERY strong profile. Be very careful about how you write your statement of purpose though, I know of people totally bombing that part of the application...

I think journals are fine (from what I gather they're actually better than conferences). Recommendations, as long as they are strong recommendations (and I gather they would be) are so much more important if they come from well-known people in the field. People in the US know to write recommendations that would get you in.

I think that, as long as you have a few safeties you would love to go to if worse comes to worst (top 50 would probably be fine), you will be more than OK.

And be sure to score decently on the GREs. By decently I mean more than 700 (I'd say more than 750) on the quant, and more than 4 on the essay. Lower scores may draw question marks, though I doubt you'll have anything to worry about it.

Good luck!

[anacron]

On 4/24/2011 at 8:35 PM, adinutzyc said:

You have a VERY strong profile. Be very careful about how you write your statement of purpose though, I know of people totally bombing that part of the application...

I think journals are fine (from what I gather they're actually better than conferences). Recommendations, as long as they are strong recommendations (and I gather they would be) are so much more important if they come from well-known people in the field. People in the US know to write recommendations that would get you in.

I think that, as long as you have a few safeties you would love to go to if worse comes to worst (top 50 would probably be fine), you will be more than OK.

And be sure to score decently on the GREs. By decently I mean more than 700 (I'd say more than 750) on the quant, and more than 4 on the essay. Lower scores may draw question marks, though I doubt you'll have anything to worry about it.

Good luck!

Thanks a lot adinutz!

I'll keep that in mind regarding the SOPs and GREs.

Congrats on the WashU admit and good luck with the PhD.

this is probably off topic, but, are there people in the forum doing Cloud/HPC? I would love to hear which colleges they shortlisted and perhaps talk about some research topics

[herring]

I'm curious about what role you believe HPC will play in the cloud (being limited by communication costs), if that's what you mean by 'hpc/cloud'

I work in HPC, and the schools that are most active / have the most faculty in the area seem to be UIUC, Berkeley, GT, UT-K, and then to various degrees UCSD,Rice(PL),UTA,Minnesota,

Don't know nearly as much about clouds but I know UW, Berkeley are good, UIUC is probably good, and GT has stuff going on.

[anacron]

I'm curious about what role you believe HPC will play in the cloud (being limited by communication costs), if that's what you mean by 'hpc/cloud'

That is exactly the topic I will be addressing with my senior thesis actually. Whether there is an appreciable difference in communication costs on a virtualized environment versus dedicated high performance clusters. (i.e Supercomputers). With network speeds increasing (infiniband, Gigabits/s scale switches) there could be a compelling reason to use Clouds due to their high availability, low maintenance and competitive costs. Gauging the performance of HPC applications can help to highlight problems and potential fixes. There's a lot of potential with clouds and HPC should be the ideal target to measure the performance of a cloud.

[herring]

So just some issues worth considering:

Cloud computing seems to be geared towards data parallelism. It seems to excel in the same areas that frameworks like MapReduce excel in: more or less embarrassingly parallel tasks. When it comes to task parallelism, you have the problem that you often can't predict how parallel a given machine is, and you haven't tuned for that machine.

Load balancing: if you have a code that executes in phases with dependencies, and some nodes become arbitrarily busy due to other users, things will start to stall. (Solution: schedulers, either before execution, or during execution (i.e, work stealing, which has many ideal asymptotic properties)).

Collective communications: many HPC codes use MPI collectives such as all-to-all. These propagate information over the entire network, and are only really efficient on certain topologies with certain locality guarantees. It would be up to the cloud vendor to specify MPI primitives and node scheduling rules to optimize for this, and that would be a very difficult undertaking on its own.

Just as a simple example, consider a parallel algorithm that expects nodes to communicate with their neighbors over and over (i.e, Cannon's Algorithm for matrix multiplication). Unless the cloud guaranteed you a set of nodes that were all connected to each other with a homogeneous topology, this algorithm would stress the network with congestion way out of proportion with what it would need on a simple grid cluster.

[newms]

As others have said, it seems that you have a strong profile and that you should have a good chance of getting an admit at most top 20 CS programs. The quality of the publication you're working on as well as how well you fit with the research at the particular school will be particularly important for your chances at top 5 or top 10 schools. Keep in mind though, that top 5 schools are ridiculously competitive, so many applicants with strong profiles (and even some with very strong profiles) get rejected. Good luck!

[anacron]

Cloud computing seems to be geared towards data parallelism. It seems to excel in the same areas that frameworks like MapReduce excel in: more or less embarrassingly parallel tasks. When it comes to task parallelism, you have the problem that you often can't predict how parallel a given machine is, and you haven't tuned for that machine.

Load balancing: if you have a code that executes in phases with dependencies, and some nodes become arbitrarily busy due to other users, things will start to stall. (Solution: schedulers, either before execution, or during execution (i.e, work stealing, which has many ideal asymptotic properties)).

Collective communications: many HPC codes use MPI collectives such as all-to-all. These propagate information over the entire network, and are only really efficient on certain topologies with certain locality guarantees. It would be up to the cloud vendor to specify MPI primitives and node scheduling rules to optimize for this, and that would be a very difficult undertaking on its own.

Just as a simple example, consider a parallel algorithm that expects nodes to communicate with their neighbors over and over (i.e, Cannon's Algorithm for matrix multiplication). Unless the cloud guaranteed you a set of nodes that were all connected to each other with a homogeneous topology, this algorithm would stress the network with congestion way out of proportion with what it would need on a simple grid cluster.

Those are some very good points that you brought up and ones we are taking into account with our model. There are certainly disadvantages in using a cloud especially with the network being a bottleneck and the overhead of maintaining the virtualized infrastructure (via Open Nebula, etc.) However, with specialized supercomputers you have the big problem of scalability, even though it offers higher throughput there is a drawback to how many CPU's, processes, etc. that are available. If the supercomputer is simply for your own use, then yes, there would be no justification for a cloud. However, if you consider various applications such as Blast, GROMACS, MM5 which can (and are) used by a big sector of the scientific community then scalability and availability are big problems. Supercomputers have long waiting lists and you often cannot use its entire computing power. This is another reason why I wanted to get into Cloud b/c there are a lot of challenging problems (scheduling, networking, load-balancing, federation). I'm curious, what are your research interests in HPC?

What I said, on second thought, might be a little confusing since I talk about supercomputers and not grids. Maybe it would be more accurate to say that the research I'm doing is geared towards seeing if HPC applications are affected by virtualization technologies, some grids do leverage virtualization which is why I did not want to generalize.

As others have said, it seems that you have a strong profile and that you should have a good chance of getting an admit at most top 20 CS programs. The quality of the publication you're working on as well as how well you fit with the research at the particular school will be particularly important for your chances at top 5 or top 10 schools. Keep in mind though, that top 5 schools are ridiculously competitive, so many applicants with strong profiles (and even some with very strong profiles) get rejected. Good luck!

Thanks, newms! You're right about the research, I really should target schools known for their presence in Cloud and HPC since I know what I'm really interested in.

Edited April 26, 2011 by anacron

[BKMD]

The quality of the publication you're working on as well as how well you fit with the research at the particular school will be particularly important

This. You'll almost certainly get into some top 20 schools with your credentials, but the quality of your publication will help distinguish you from other applicants at the most competitive (top 4) schools. It's increasingly common for undergrads to have publications under their belt, so this isn't an automatic advantage, so where you really have a chance to shine is by not only publishing a paper but publishing an idea that's non-trivial and actually moves the field forward, rather than an obvious idea with straightforward results. It's good that you're the one who came up with the idea - this is important and your letter writers will mention it.

And of course fit with research interests is another major factor - when professors are deciding which prospective students they are interested in advising, a match in interests is very important. You probably won't get admitted even with the best credentials if the professors don't think your interests would align what theirs', and it's amazing how many applicants don't realize this. So my point is not to stretch yourself thin and apply to too many places since many of them won't be a good fit anyway, but focus in on schools that you really are enthusiastic about with professors who are already doing what you want to do. Play up this fact when you write your essays, and be specific about who you want to work with and why. Also keep in mind that the best professors in your particular area of interest might not necessarily be at schools with the highest US News ranking. A lot of applicants ignore this or don't realize this and it's a shame.

I think you'll have a good application, so good luck!

[anacron]

Thanks for the great advice BKMD. You're definitely right about looking for prospective advisors instead of just the school rankings. On a related note, do you guys think its good to get into contact with potential advisors - now? The reason I ask is, I will be at UIUC for the summer and there are a few professors who are active in my area of research. So it would be a great opportunity to chat and see if our interests are similar.

[Amogh]

Thanks for the great advice BKMD. You're definitely right about looking for prospective advisors instead of just the school rankings. On a related note, do you guys think its good to get into contact with potential advisors - now? The reason I ask is, I will be at UIUC for the summer and there are a few professors who are active in my area of research. So it would be a great opportunity to chat and see if our interests are similar.

Do that.. it's always good to network. and now that admissions season has come to an end profs might just be more willing to talk to you. and giving them a heads up that you will be there at some time in the future is a good way of showing that you'r interested in their work. just make sure to mail them all your final plans just before you set off so they can make some time to talk to you.

Nice profile btw.

[anacron]

Do that.. it's always good to network. and now that admissions season has come to an end profs might just be more willing to talk to you. and giving them a heads up that you will be there at some time in the future is a good way of showing that you'r interested in their work. just make sure to mail them all your final plans just before you set off so they can make some time to talk to you.

Nice profile btw.

Thanks Amogh and thanks everyone for the great advice.