This test has six questions, all of which should be answered. The test runs the entire class period, starting at 7:25 and ending at 9:15. When you finish the test, you may leave.
Always justify answers unless the question explicitly states it’s not necessary. Some questions require reasonable assumptions in order to be answered. In such cases the answer should clearly state the assumptions made.• Sustained bandwidth• Programmed I-O• Block-level striping• The Sockets interface• Boot block• Interval timer
• Sustained bandwidth:The steady-state bandwidth (transfer rate) achieved during an extended I-O operation.• Programmed I-O:CPU-specific• Block-level striping:Spreading contiguous data across several disks by breaking the data into blocks and writing successive data blocks to the same disk block on successive disks.• The Sockets interface:An operating-system supported interface to the networking subsystem.• Boot block:A disk sector containing code that is read in to primary store and executed as part of system boot-up.• Interval timer:A register that receives some number of ticks, decreases the number periodically, and triggers an interrupt when the number reaches 0.
The answers, approximately verbatim, and in no particular order:
• Sustained bandwidth:The amount of data the device transfer divided by the amount of time the transfer took.• Programmed I-O:A type of communication method that uses the DMA to communicate across devices in order to avoid wasting CPU with interrupts.1• Block-level striping:A type of parallel data storage/retrieval method that spreads sequential blocks of data across multiple hard[?] disks.• The Sockets interface:An interface used for network communication that drivers can use to obfuscate complicated network actions.• Boot block:The section of the hard disk/secondary storage that contains the code for loading in the operating system.• Interval timer:A register used for timing.2
1: Mostly the opposite, but without DMA.
2: Used how?
• Sustained bandwidth:When the flow of data remains at a constant speed.• Programmed I-O:The information3 that passes4 between the system and storage device that is used by the system.• Block-level striping:Removing small sections of data from a disk at the block level.5• The Sockets interface:The connections6 made between the system and the storage device.• Boot block:The section of the disk that first gets loaded.• Interval timer:A mechanism that records the amount of time between device I-O operations.
3: Any particular kind of information, or just information in general?
4: Passes in a particular direction?
5: What happens to the removed data? Where does it go? Why is it removed?
6: What kind of connections?
• Sustained bandwidth:Open channel of data flow staying at a specific average, little variation.• Programmed I-O:[ no answer ]• Block-level striping:Place pages across entire blocks.• The Sockets interface:[ no answer ]• Boot block:[ no answer ]• Interval timer:[ no answer ]
• Sustained bandwidth:Refers to the amount of data that can be sent/received/written retrieved at a time.7• Programmed I-O:Hardcoded commands8 for the CPU/system to start;9 simulated input to receive identical output result.10• Block-level striping:Designation of a sector before and around the data11 for future data of a similar nature.12• The Sockets interface:Polling on a socket13 for specific information to be sent to and from a process.• Boot block:Starting section of disk holding all data necessary to get the core of the system running. Allows OS to be “up” prior to loading extraneous drivers, data.• Interval timer:Timer based off systemCLK that is usually used in a polling scenario, to check for something (ex: keypress) in the last interval (ex: 1000 ms).
7: Over many times, hence “sustained”.
8: Hardcoded where?
9: To start what?
10: Output result identical to what?
11: Which data? Where is the data?
12: Data similarity determined how?
13: If a definition contains the phrase it's trying to define, it's probably not a good definition.
• Sustained bandwidth:Consistent14 throughput between device to computer connection.• Programmed I-O:The part of the OS that handles devices connected to the computer, and how they interact with the OS.15• Block-level striping:Block i in a sequence gets stored in disk i mod n. Reduces per-disk queuing.16• The Sockets interface:[ no answer ]• Boot block:[ no answer ]• Interval timer:[ no answer ]
14: Not the best word to use; “long-term average” is more to the point.
15: Way too broad a definition.
16: Perhaps, unless the request is in start+length form.
Disk striping breaks up a contiguous data sequence and stores the pieces on a set of disks. Disk striping's chief benefit is parallelism: it turns a single large disk I-O operation into a set of n concurrent I-O operations, effectively increasing the transfer bandwidth by n when there are n disks in the set. The costs are the cost of the extra disks and the increased complexity of read and write operations.
The answers, approximately verbatim, and in no particular order:
1: Maybe marginally more expensive, but not fundamentally so unless the sizes are extreme.
2: But doesn't RAID at its own costs?
3: Wouldn't parallel I-O increase performance in most cases?
4: Possibly true; what examples would illustrate the claim?
5: Doesn't sound practical in that case.
6: A multiplicative factor, which combine to an exponential increase under independence assumptions.
7: Which this?
8: Nope, not even close. This reads like some quasi buddy-allocation policy.
The DMA C would have to map virtual address to physical addresses and pin the virtual-physical mapping information so the storage-management subsystem won't try to page out the otherwise unused pages.
The answers, approximately verbatim, and in no particular order:
1: Any particular kind of controller?
2: Any other coordination required?
AA1016 in the source pointer and the destination pointer.3 In order to accommodate the change to
virtual addresses, the pointer must be changed to also include the offset of the
function,4 so the same length pointer5 would now be [ diagram ]. The physical
implementation of “number of bytes to transfer” would also have to
change, as the bytes may not be sequential, instead spread across the memory in
different settings with other programs in between.6
3: Is the single memory address virtual or physical?
4: Which function?
5: Length pointer? What's that?
6: Is non-sequential layout in page frames something that's visible at the virtual-address level?
The most important characterization is CPU utilization. If utilization is low, CPU cycles are cheap to free, depending on how low, and spending them on polling would not be a problem. Otherwise CPU cycles are expensive because of high demand, and interrupt-driven I-O is more cost-effective than is polling.
The answers, approximately verbatim, and in no particular order:
1: “Periodically” is a more accurate description.
2: Or when the device is done performing an operation.
3: Or whatever.
4: But what if your colleague's system doesn't have much CPU demand otherwise?
5: Or periodically.
6: Oh? Easier than polling? Really?
7: As is true of anything.
8: By what performance measure?
9: I-O is usually done only when requested, although the requests can come from outside the system.
10: Would this suggestion hold under all circumstances?
Speed of system reaction: Immediate vs soon.11
Say HW is important, but not critical, with fast-as-possible responses.12 Pick polling with a high rate; this allows system to not drop concurrent actions and react to HW input/action at a fast speed.11: Some kind of measure of latency, yes?
12: Responses to what? By what?
Seek time and rotational latency are the two main components of response-time variance. Seek time costs more in time than does rotational latency, and disk-arm schedulers have almost no influence over rotational latency. Under these assumptions, seek time is the most important component of the response-time variance.
Because FCFS scheduling minimizes disk-arm travel, it should have smaller response-time variance among the disk-arm schedulers. FCFS can cause starvation, but that will effect service time, not response time.
The answers, approximately verbatim, and in no particular order:
1: The question asks about the variance in response time, not wait time.
2: One end of what to the other?
3: All the requests? Or maybe just some of them having some property?
4: The question isn't about starvation.
5: But perhaps it is because of starvation that SSTN may have a low response-time variance.
6: But perhaps bounded time-to-service?
SCAN: max addressable memory locations x + x2, assuming the disk head starts at address 1 and needs to read address 0 with a velocity to the center of the disk.10
C-SCAN: max addressable memory locations11 x + + + speed to 0 address, assuming the same as above.LOOK: equivalent to SCAN.
SSTF: ∞, assuming that incoming addresses are always smaller12 distributed from the head through the requested address.FCFS: (max number of requests in queue - 1) + (time for head to move from address 0 to max addressable) x t,13 assuming the request is at the end of the systems ability to schedule addresses.14
FCFS may have a lower variance than C-SCAN, but unfortunately that is dependent on the ratio of schedule address queue to disk size.157: This is a terrible assumption because neither the maximum nor average variance is a quantity that makes sense on anything other than an arithmetic level.
8: How does this statement make sense? Different disk-arm schedulers have different response times, so how can they all have the same minimum?
9: What does this mean?
10: What is being described by this sentence?
11: What are “max addressable memory locations”? What are they measuring? Why?
12: Incoming addresses are smaller than what? And what's an incoming address?
13: What is this equation describing? How does it make sense to add a queue size and a time?
14: What does it mean for a request to be “at the end of the systems ability to schedule addresses”?
15: How does such a ratio have any meaning?
16: Which ones?
17: Hard to imagine a useful activity that can function without any kind of input.
18: A vertical slice of the disk? What's that? A cylinder? And don't all disk-arm schedulers use a vertical slice of the disk?
19: Faster is not what the question's asking about.
An unending sequence of I-O requests for sectors in the same track can cause all disk-arm schedulers except FCFS to starve processes issuing requests for other tracks. Rotational starvation can be dealt with the same way seek starvation was: limit I-O behavior per track by allowing only sector operation (read or write) per track visit, or ordering the sector operations so that once an operation on sector i has been performed, only operations on sectors j > i can be performed.
The answers, approximately verbatim, and in no particular order:
1: Is this answer confusing starvation with deadlock?
2: What process?
3: But wouldn't the restarted process, whichever process that may be, just go back to the previous circumstances in which starvation was occurring?
4: But priority can cause starvation too.
5: “Intelligent” means what?
6: Proximity of what to the disk arm?
7: Is the priority dynamic? Suppose a process has several requests in the queue (possible with asynchronous I-O); what then?
Or a c-scan scheduling with the two arms going in opposite directions.9 More plausibly, smaller disks, but even more of them, reduces the length in which the arm must travel. There would be less room for data but the rate to retrieve that data would be extremely fast.10
More realistically11 you could have the data of one disk duplicated to another and they would have two different schedulers, or c-scan in opposite directions, however the overhead on it would be too costly since every bit would have to be written twice with the same address.128: In half which way: radially or circumferentially?
9: What?
10: Does this paragraph have anything to do with starvation?
11: ?
12: But wouldn't the bits be written in concurrently?