Should CS students be graded on performance? | Hacker News Reader

Should CS students be graded on performance?

sbcoded.com

22 pointsdrfugly34 comments

It depends on the context of the problem:

* If it's being framed as being applied to a small number of items on one machine, I'd mark down for premature optimization. Normally such assignments have a specific point, and going off the deep end with efficiency measuring often misses it.

* If it's being framed as being applied to huge data sets, asymptotic complexity is the metric to go off of.

* If it's being framed as a latency-dependent (UI, servers) then actual timing is the metric to grade off of.

I'd say that if you have an algo that is faster for small numbers, it usually doesn't really matter that it's faster. Small numbers go fast no matter what. Unless it's latency dependent, it's encouraging students to work on things that don't matter at all.

EDIT: spacing EDIT2: spelling

When you were in class, where you ever even introduced to the concept that sometimes it might be possible that O() notation might not tell the whole truth? I understand that for the most part you shouldn't have to worry about it, but shouldn't we still warn students about this?

Do you mean something like the case of quicksort mentioned in the post?[1] I was taught average- and best-case analysis along with worst-case analysis. Or do you mean some other fact not revealed by that analysis?

[1] "Quicksort is a magical algorithm that theory tells us runs in O(n^2)"

Right, but there is also the cases where a nlg(n) quicksort for the most part is slower than the O(n^2) version in the real world. I wonder if current CS teaching is completely ignoring real-word performance.

Could you explain what you mean a bit more by big O notation not telling the whole truth? Are you referring to the fact that algorithms that look the same in big O notation might have drastically different constants, which can affect real-world performance?

Yes, that's exactly what I'm getting at. Or cases where multithreading an application may look like it's going to bring obvious performance gain only to see the overhead kill your performance.

It can be worse. A O(n log n) algorithm might be slower than a O(n^2) algorithm for all practical values of n.

Also, one should be careful what to count. Sorting strings, for example, is not quite O(n log n); average string length/expected offset of first difference/whatever should also be in that O().

Along the same line, for many algorithms, cache-locality is more important than number of CPU cycles. So, counting cache misses rather tha cycles can be the better way to judge an algorithm.

Absolutely! I suppose when I speak of asymptotic analysis, I'm including average case/expected case/degenerate case analysis in the whole exercise. Perhaps "runtime analysis" is a better term - it can then incorporate latency analysis as well.

But yes, that was integal to the treatment of the subject. I remember having to determine what the worst-case of quicksort looks like as a part of an assignment to exhibit, in practice, best/average/worst case runtime of a number of sorts - and this was an freshman-level intro course. That particular problem was one of the most fun homeworks I've ever had - a rather satisfying solution.

We talked about these things as well. I remember studying how to calculate the percentage of data sets of a given size that would bring out the worst-case running time of quick-sort.

Really, I think the issue with all the criticisms of asymptotic analysis is simply that too many people (even brilliant programmers and CS majors) just don't understand what big O notation actually means. If an algorithm is in, say, O(n lg n), that says nothing about how fast it runs with 10 inputs, 1,000 inputs, 1,000,000,000, etc. It merely says how its running time changes as its input size increases. The algorithm could literally take 1,000 years with an input size of 10. That doesn't matter. At some input size, it will run faster than a different algorithm in O(n^2) that completes in 1 millisecond with an input size of 10.

The algorithm could literally take 1,000 years with an input size of 10. That doesn't matter. At some input size, it will run faster than a different algorithm in O(n^2) that completes in 1 millisecond with an input size of 10.

𝜪 is an indicator of work (number of steps) for a given n, not wall-clock time. This gets mildly confusing when you give each unit of work a value of 1 unit of time, and then talk about it in terms of time-like labels (seconds, hours, age of the universe).

I don't see that there's value is comparing an algorithm of n = "some input size" to another one with a different 𝜪 of n = 10. When comparing, you don't care about the value of n, you only care about how n changes the amount of work. When actually selecting and implementing, you care about n (because n will often be limited by something else, say available memory) -- if your n is small and pragmatically you know that even a terrible, brute-force algorithm will finish in a second, you use the one that is easier to implement and put an implementation specific limit on n (and you also put a TODO or FIXME on it with a comment that says if n ever needs to be increased, a different algorithm should be used).

Yes, a few professors discussed situations in which a theoretically slower algorithm was better in almost all real situations, and situations like timsort in which it is best to fall back on an asymptotically slow algorithm when your inputs are very small.

We were taught what big O notation actually is, so yes. The whole point of big O notation, what makes it so useful, is that it hides information (namely constant factors). It's not supposed to "tell the whole truth," it's supposed to allow the analysis of algorithms not at any one runtime, but in an asymptotic mindset.

With big O notation, you're not interested in how fast an algorithm ever actually runs (for that is the realm of constant factors), but rather how its running time changes as the size of its input changes.

Actually, "fast on small n" does matter. Many real-world qsort() implementations have an insertion sort in the inner loop (e.g. for sorting chunks of 8 elements) - and while the number of elements is small, this routine is called often enough that it matters.

That said, some abstraction is useful too.

But this optimisation is applied many times: it is the leaf case of such qsort algorithms. So it affects the constant scaling factor at all ns, not just for small ns.

I thought I was being totally clear, but seeing your comment and its upvotes, apparently not. What exactly did your "but" refer to? I think we're in complete agreement...

(I was trying to say: here's an example where properly using a fast-for-few-elements "toy" algorithm can improve real-world performance on a "serious" problem like sorting a big chunk of data. So big-O is not everything.)

The point is that this improvement doesn't deliver a speed up that is significant only for small n, but a speed up everywhere.

I agree that fast on small n does matter, it's just that this isn't an example.

Thanks for clarifying! (Still, you get a global speedup because the "inner" algorithm is fast for small n.)

Aside from teaching big O notation, I would be skeptical of grading for performance. Teaching people that the performance code they write to solve a toy problem matters might lead them to go into industry and then perform microoptimizations on code to solve performance problems when they are virtually universally caused by the interaction of many moving parts.

If you're doing web apps, application code is close to the last place I'd reasonably expect there to be severe performance issues. You're far, far more likely to either blow something architecturally ("We call a foreign API in the request/response cycle.") or flub settings where simple best practices produce repeatable results ("It is 2011 -- do you know where your gzip is?")

First, find a better color scheme for something you intend people to read. That was painful.

Second, when I was a CS student I expected to be graded on my response to a particular problem. If you wanted me to roll my own sort, I expect that to be specified. If you want me to explain in comments why I chose certain algorithms say something in class. I'm not going to waste time implementing an efficient sorting algorithm for a 100 element list (or 1000 even) just to provide a TA with a nerdgasm.

I feel this article is asking if students should be graded on an arbitrarily deep rubric. I had enough to worry about as an undergrad without caring if a bored TA was going to dock me 10 points for calling Arrays.sort.

Will do for the color scheme.

What I was trying to ask was: Should real world performance be more heavily weighed? I feel like right now all the CS curriculum focuses on is O() performance as the end all.

Normally, my (favorite) CS professor would weigh code clarity more strongly than performance. However, he would sometimes put up assignments that focused almost entirely on the performance, in which he obviously changed the grading weights.

First, find a better color scheme for something you intend people to read. That was painful.

http://rdd.me/3ltgecrd

On my iPhone safari, readability killed all the big Os, makes the article very hard to follow.

I tried for a moment to figure out how to deselect all the text.

YES. My Data Structures and Algorithms class was graded by benchmark (ECS 60, Sean Davis, UC Davis). Being graded this way was one of the most valuable collegiate experiences I've had (I'm a fourth year on the way out).

Sean gives his students two "challenge programs" a quarter. The first program is usually ultimately a matter of picking/optimizing the right data structure. We do graph algorithms by the 2nd half and so the later challenge problem is inevitably some graph algorithm.

The challenges are close to real world problems. For ex., our 2nd challenge program was to determine the maximum # of donations Obama could receive at any one time given a capacitance graph of his website's network topology (I took the class around the '08 election). We had to design, implement, and test everything by ourselves.

I don't agree with the naysayers who suggest that grading this way doesn't teach the really important software engineering tasks. You really had to feel out the problem domain in order to get the best optimization, and it the book answer wasn't always the best answer. Optimization is about a lot more than simply picking the right data structure or algorithm.

Partnering with the crazy russian genius with an 8086 instruction set in his head that quarter made me the C programmer I am today.

TL;DR Grading by benchmark makes students better programmers.

The ACM contest has it right: specify the size of the input and the running time available. The program must be able to run an input up to that size in the given runtime. That's it: no more, no less.

At Harvey Mudd College, we have a class based on the ACM which works similarly, though with relaxed runtime constraints: 42 minutes per problem. In practice, you rarely need more than a few seconds, but the 42 minute cutoff (combined with large problem sizes) is a great way to prevent people from using brute-force approaches when there's a smarter option available.

42 minutes? I recall more than one ACM problem that could be bruteforced just within the running time (in C). Actually, I did really badly in one contest because I applied serious mathematics to one such problem (which gave a much faster program, but writing it took a lot of time and fast enough is fast enough.)

Sure, some can be bruteforced, but typically it's intentional.

A good example of what I mentioned was an assignment this week. Last week, the input size was constrained to a thousand or so, and Floyd-Warshall was able to solve the problem. This week, it's the same problem, but with inputs up to ~50,000 in size. The old program, run on the new test cases, doesn't finish in a day.

No, they should not. Unless performance was one of the assignment requirements.

Performance is rarely on the most important requirement in software development.

If you don't believe me, watch Prof. Leiserson make this point.

http://ocw.mit.edu/courses/electrical-engineering-and-comput...

Define performance. The question is ill formed since there is nothing specified to measure. The example wonders who should get the better grade based upon design decisions each of which are defensible in some context. And it's the asymptotic and/or typical behavior that is touted, not the implementation.

At a most elemental level, a program written by a CS student should work, that is, do what it is supposed to do for all legal inputs and provide meaningful diagnostics when it fails. That's hard, but that is the nature of programming.

Finding and choosing algorithms to accomplish a task (well defined or not) is also computer science. That is algorithmics and design. Doing a good job here is harder to measure and depends both a deep understanding of the available building blocks and the problem at hand. Moreover, trade-offs are involved and need to be resolved even though there is little real data to guide decisions. Seeing and exploiting invariant properties of the problem often enables better performance in some dimension; insight there is often called cleverness.

Equally important is the ability to see complex things in a simplified way by introducing abstractions. That's the only way to solve difficult, complex problems, and it is exceptionally difficult to codify.

I had a CS Professor who was convinced that he had cheaters in the class(1), so he decided we should all be graded on performance of our programs. This resulted in a lot of stress for a lot of students (C code only, most students taught only Pascal/Modula-2) and some seriously nasty things done in the programs. I rate it a big fail for what we were doing.

I can actually kind of see this for embedded / small device programming. Embedded systems that can't meet their performance goals are useless and buying faster hardware is very costly option.

(1) he also believed not one of us submitted a C program to print "Hello, World!" that worked. It was a 8AM class and I had just driven 120miles (home visiting) through a blizzard to make it to class when he announced this fact in a dramatic fashion. I required calming down.

How would grading on performance discourage cheating?

Generally, performance isn't interesting in cases where asymptotic running time isn't dominant. You could sort 3 items with a couple of `if` statements, which would compile to a handful of microprocessor instructions. That's not interesting. Surely, in any data structure or algorithms course offered as part of a CS degree, the sorting assignments are for the "general sorting" problem. This implies relatively large data sets (and unpredictable, since you don't expect to be able to use a faster non-comparison sorting algorithm).

No. Stuff like this only comes with experience and it's more important to teach students about the fundamental and overarching themes of computer science than to pester them with optimization issues.

Reply on news.ycombinator.com