University Stanford Computer Science Department CS 240 Quiz 1 Spring 2006 May 12, 2006
This is an open-book exam. You have 50 minutes to answer eight out of ten questions. Write all of your answers directly on the paper. Make your answers as concise as possible. Sentence fragments ok. NOTE: We will take o points if a correct answer also includes incorrect or irrelevant information. (I.e., dont put in everything you know in hopes of saying the correct buzzword.)
Question Score 1-5 6 - 10 total
Stanford University Honor Code In accordance with both the letter and the spirit of the Honor Code, I did not cheat on this exam nor will I assist someone else cheating. Name and Stanford ID:
Signature:
1
Answer eight of the following ten questions and, in a sentence or two, say why your answer holds. (5 points each). 1. Given a program P that runs on Caprpriccio and does no explicit locking of its own. What does P correspond to in terms of MESA? What is the most straightforward way you could adapt Eraser to nd race conditions in such programs? (Hint: think of how it checked OS code.) How useful do you expect this to be? P = one big monitor with one implicit condition variable, yield = wait statements on this variable. Since there are no locks, Eraser wouldnt be able to check much. You could have one lock for the monitor that was implicitly acquired when a thread starts running and released at yield point. However, this isnt very useful since it would be held whenever P ran. Finding races in this case would actually require looking for violations of invariants, which Eraser is not setup to do. Grading: (+2) saying that the program was one big monitor with words implying that this gave a form of mutual exclusion through non-preemption or that whatever thread was running had exclusive access, etc., (+1) by tying the wait and notify to thread yielding and scheduling and (+2) for identifying that Eraser wouldnt be useful because there are no locks (except maybe the one big monitor lock shared by everyone) in the program, so hence the lockset algorithm is useless.
2. Consider the Therac threading model. How does it compare to Events in Ousterhout? To threads in Capriccio? Does it contradict or support the arguments of these people? Therac is all screwed up, so we had to take a variety of answers. In some sense you can see it being event-driven, since there is a sort-of event loop that decides what thread to run. In another sense it is standard pre-emptive thread with the standard default (nothing is a critical section). It can be seen as a vindication of the pre-emptive threads are really hard mantra. But programmer would have probably screwed up any model, so things are a bit unclear... Grading roughly: (+2) to make the connection between events and how Therac dispatches certain tasks, (+2) observing that the threads in Cappricio are non-preemptive (not preemptive!) and has nothing to do with the preemption problems in Therac (although talking about implicit yields was ok), (+1) making a lucid argument why this does or doesnt support Ousterhouts claim; the main issue here was confusing Cappricios threads with pre-emptive threads, since the claim of the authors of the Cappricio paper is that you get the benets of events (we avoid the problems of pre-emption) without the fuss of continuations.
3. Consider very-broken code that uses the double-check lock idiom:
0: int *p = 0;
2
.... 1: if(!p) 2: lock(l); 3: if(!p) { 4: int *t = malloc(sizeof *t); 5: *t = 3; 6: p = t; 7: } 8: unlock(l); 9: } 10: x = x / *p;
How can an optimizing compiler of the sort that Boehm complains about interact badly with the use of the t temporary? Assume you dont use an optimizing compiler: if another thread has done a free previously, what can happen if the initializing thread gets context switched immediately after line 6 and another thread executes line 10 on another processor? It could remove the temporary assignment and assign directly to p, which means that another thread could see that p was not null before its contents were initialized. If the free had happened on another processor, then that processor could still have a stale copy of that memory in its cache. If this memory was then reused by the malloc and assigned to p and the second thread ran on the processor with the stale data it could see the out-of-date copy. This would be hard to debug. (Note, in reality that malloc and free probably use locks which will probably cause memory synchronization, which would prevent this bug from showing up. Grading: roughly: (+2.5, rounded up with the total sum) talking about how an optimizer could screw things up; main example was optimizing away t, but other another example included reording the statements in the critical section (a problem since the access to p on line 10 is not protected by a lock). Some people said that the compiler could move the call to malloc outside of the critical section; for a general function this is completely false; because of potential side-eects the compiler cannot reorder function calls unless it can perform a whole-program analysis and determine that functions can be reordered because one does not eect the other. Because malloc is often an intrinsic function, the compiler can place the call to malloc outside the critical section, but the assignment to p must be done within the critical section (we dont believe, however, that a compiler like gcc does this in practice). Another possibility is that the inner if statement could be optimized away, causing issues. This is potentially possible since p is not declared volatile. A variety of other answers were closely evaluated... (+2.5, up rounded with the total sum) the main issue with the call to free is that the pointer value can be stale across processors (for many reasons). Other answers were also considered on a case by case basis. Partial credit was given (removed) for technical points that were unclear or incorrect.) 4. Your 240 cohort sees the number 98.2 in the VMS paper and says that is strange. Why? Assume the methodology that obtained it was accurate: What is a plausible 3
reason the number is what it is? This is the average number of pages written. They claim the high-low limit is around 100. One plausible explanation is that pages with pending I/O operations were written and then rewritten when the I/O ops completed. Grading roughly: (+2) that it was clear that 98.2 was the number of pages written. Some answers would conate this number with the number of pages read, or simply did not clearly indicate they understood what this number was. (+1) about some discussion of the modies list, or something to the eect of where the pages came from and that they were batched, or a reasonable explanation of why the number was around 100 (+2) a exible answer to identify why the number is 98.2 as opposed to something like 100 (reasoning in an experimental way about why the number was the way it was).
5. Superpages: Jimbob says the most important thing to know about the future is when an object will die. Bobjim says that the most important thing is how an object will be populated. How could you use these facts? Give an experimental reason for why one of these guys is more right. You could use death times for placement: put things that will die together near each other so that death frees up a lot of space. You could use population information to decide on whether to do eager promotion or not. From the experiments, the overhead of lazy promotion and general management doesnt seem to be such a big deal, so probably placement is more important (since it lets you use superpages when memory contention is much higher and the benet of superpages for many apps is non-trivial.). Grading: (+2) to reason about how Jimbobs policy can be used to eectively reclaim pages, etc., especially in the context of how this helps management of superpages. Many answers did not talk about how this was useful from a page management perspective (this question is about superpages!). (+2) to reason about Bobjims policy, and how it could be used for eager promotion, or something useful for page management (again many answers ignored the eect of the policy on superpages). (+1) a reasonable argument, one way or another, of why one policy was better than the other by comparing the benets of both as put in the previous two points. Some answers simply said that one was better than the other without mentioning what the benets of one or both facts were. Both facts are potentially useful, and to get full credit you needed to discuss both of them.
6. Most OSes dont have superpages. VMware hires Navarro to hack ESX so that ESX will trick guest OSes into using superpages. At a high-level: how would he do so? Do you expect this approach to work well? Why or why not? What he would want to do is to map the PPNs of contiguous VPNs to contiguous 4
MPNs and insert this into shadow page tables to they get put in the TLB at miss time. However, this probably wont be that eective: the guest OS will just pick PPNs it wants, which...