Sample Quiz 1 Questions/Answers

For the quiz, you may bring the class notes, the papers that were assigned, your writing assignments, the presentation you prepared, and the text book (assuming it arrives...)

In Ethernet: Distributed Packet Switching for Local Computer Networks, the author describes a link-layer protocol used to send packets around a local computer network. The physical medium for this network is a coax cable (like the kind you get cable TV from). The communicating computers are connected to the same cable at the same time.

Q1: When two computers try to inject a message into the cable at the same time, the messages collide. What does the Ethernet Protocol say the computers should do to try to get their respective packets out?


Each computer waits before trying again, using a Binary Exponential Backoff algorithm. The wait time is determined by computing a random number then multiplying it by an estimate of the network load. Subseqent collisions lead the algorithm to exponentially increase the waiting time.

Q2: You are getting bad performance on your local Ethernet network. A little investigation shows that your nemesis, Net Vader, is watching lots of videos - and he's sharing the same Ethernet cable as you! To retaliate, you decide to create a rogue Ethernet card that gives preference to your packets over NV's during a collision. What will your card do different when a collision is detected? Be concrete.



Any of the following gets full credit - this isn't a complete set of answers

For a collision that results when trying to send data, the rogue ethernet card sets its waiting time to zero. This gives the rogue card's data a higher probability of making it out.

Before sending a request, the rogue card can be used to disrupt any video streams by creating collisions that cause NV's card to back off exponentially.

The rogue card can create collisions whenever it sees packets addressed to Net Vader's machine. Hopefully, this will cause NV to give up watching videos altogether.

Q3: The paper describes the use of a packet transceiver for connecting two coax cables into what looks like a single cable. Describe how such a system should handle collisions. (Hint: Consider two cases: A collision on one cable but not the other, a collision on both cables at the same time)



Collisions on any point in an Ethernet cable need to incapacitate the whole cable. To act as a single cable, a collision on one of the two cables must be propagated to the other cable. Collisions on both cables at the same time need not be propagated.

Q4: Do you think the tranceiver setup described in Q3 would scale out to hundreds of cable segments connected together? Briefly describe why you think it would or would not.



Either of the following two gets full credit (there are other potential answers).

1> No. As more hosts are added to a single "virtual" cable, the steady-state load on the cable will likely grow. If the steady-state load surpasses the optimal load (as described in the paper), the performance will be severely degraded.

2> No. Even if there are few hosts, the following can occur:

As more and more physical cables are connected into a virtual cable, the propagation of collisions to the entire network can have a devistating effect on performance.

Consider a simple example - two computers (C1, C2) at the end of one physical cable, connected to two computers (C3, C4) at the other end of a connected physical cable. These cables are joined into one virutal cable.

C1-C2----------------------| packet tranceiver/repeater |----------------------C3--C4

Suppose C1 and C2 are communicating, and have a collision. That collision must be propagated to C3 and C4. Suppose C3 and C4 subsequently have a collision. That collision must be propagated back to C1 and C2. Thus, even hosts that are physically far away can break down local communication. Better to install packet filters and treat the two cables as two separate subnets instead of one virtual cable.

In Disk System Architectures for High-Performance Computing, you learned about the 5 RAID levels. Suppose you are configuring a service like photo.net with three RAIDs. In the questions below, make sure you explain your choice and list the key assumptions you used to choose the RAID you think is most appropriate.

Q5: You must first configure a RAID for photo.net's swap file. Given your knowledge of virtual memory, what RAID level is most appropriate? (assume photo.net has enough main memory to prevent thrashing)



We're assuming there's no thrashing. Let's also assume that, as stated in lecture, the virtual memory manager wants to read/write in large blocks. Let's also assume that we want to minimize read errors, as corrupted memory could bring the server down. Finally, as we heard in lecture, let's assume there's only one virtual memory manager (i.e., we need to do only one I/O at a time). Thus, we choose RAID Level III. This gives us lots of bandwidth for reads and writes, plus good error correction. While Level II would have been a good choice, Level III will give us better performance (especially if we replace the check disk with a solid-state device)

Q6 Now, you must configure a RAID for photo.net's JPG files. These are the photo files that Philip or users upload. Assume there are 50,000 JPG files on photo.net and that 250 jpg files are uploaded each day. Also assume that photo.net must serve several requests for distinct JPG files at any one time, all day long. What RAID level is most appropriate?



We're assuming that we want to access different files at the same time, and that writes are not common (e.g., ~10/hour). We're also not concerned if a few bit errors are not detected or corrected. Who's going to notice? Thus, Level I would give us good performance.

Q7 Finally, you must configure the RAID for the Oracle database files. Like the JPG case, there are many more reads than writes. Unlike the JPG case, Oracle uses only a handful of very large files. Most of the time, the same file is being read or written. What RAID level is most appropriate?



Let's assume that even though there are more reads than writes, that there will be lots more reads and writes in total than in the JPG case. This makes sense - every time users input data to the website, that's another database write. As we get more users, we'll get more interaction, leading to more writes. We also need data integrity, and improved read/write bandwidth over a single disk. Thus, Level V is most appropriate.