Networked Systems (SAMPLE QUESTIONS), COMPGZ01, May 2016

Transcription

1 Networked Systems (SAMPLE QUESTIONS), COMPGZ01, May 2016 Answer TWO questions from Part ONE on the answer booklet containing lined writing paper, and answer ALL questions in Part TWO on the multiple-choice question answer sheet. Marks for each part of each question are indicated in square brackets Calculators are permitted Part ONE 1. Ethernet, the Capture Effect, and TCP Recall that Ethernet resolves collisions using binary exponential backoff (BEB). a. Fully describe Ethernet s BEB scheme, including its effect on Ethernet s transmissions, how it behaves after both collision-free transmissions and collisions, and any limits on BEB s behavior. [5 marks] b. Suppose three hosts, X, Y, and Z, are attached to the same Ethernet network. X and Y both attempt to transmit their respective frames for the first time and collide. When X and Y contend to retry their transmissions, Y transmits next and no collision occurs. X still needs to retry its transmission. Suppose that while Y is still transmitting, Z s Ethernet interface finds that it has a new frame to send, and Z senses Y s carrier. And suppose that Y has no further frames to transmit. Which Ethernet host, X or Z, is more likely to send next after Y? What is the probability that each of X and Z will send first without a collision, and the probability that X and Z will collide? Derive your answer based on the relevant details of the Ethernet MAC protocol (including any relevant details of BEB). COMPGZ01 1 TURN OVER

2 c. As described in 3035/GZ01 Coursework 2, the capture effect occurs in Ethernet when senders X and Y collide, sender X wins in contention to retransmit (i.e., X retransmits before Y ), X then tries to send its next frame, and in doing so collides with Y s retransmission of its original frame. In this scenario, X will draw its backoff from a shorter range of durations than Y with the result that X is more likely to send a new frame than Y is to retransmit a frame that suffered a collision. This pattern is likely to repeat for subsequent collisions, such that X obtains a much greater fraction of the Ethernet s capacity than Y. Suppose that X and Y are the only two hosts on an Ethernet, and that X uses the Transmission Control Protocol (TCP) to send data to Y over this Ethernet. And suppose the capture effect occurs on this Ethernet, such that TCP sender X has far more opportunities to send packets over the Ethernet than TCP receiver Y. Assume that X s congestion window size (cwnd) has already reached a value of 64 packets, that there are 64 packets already in flight between X and Y, and that Y s advertised received window does not limit the number of packets in flight between X and Y. How will the capture effect affect the TCP transfer from X to Y? Refer to any relevant details of the TCP protocol in your answer. [3 marks] d. Your 3035/GZ01 classmate Mac Phair proposes the following change to the Ethernet MAC protocol with the intent of preventing the capture effect: Suppose an Ethernet sender X sends its first frame successfully (without a collision), then attempts to send its second frame, but detects a collision. In this specific case, X should use a modified BEB as follows. X backs off two slot times before attempting to retransmit the frame that experienced the collision. If this retransmission also collides, then before attempting its next retransmission of the same (second) frame, X backs off zero slot times before this second retransmission. If this second retransmission also collides, then X reverts to the standard BEB for this and all subsequent retransmissions. In all other cases, senders use Ethernet s usual unmodified BEB. Suppose you have a two-node Ethernet consisting of sender hosts X and Y. Suppose that both X and Y have long queues of frames to send with their Ethernet interfaces. Does your classmate s proposed change to the Ethernet MAC protocol eliminate the capture effect for such two-node networks? Why or why not? In your answer, assume that all Ethernet frames transmitted are many Ethernet slottimes in duration. [8 marks] COMPGZ01 2 CONTINUED

3 e. Suppose you have an Ethernet of more than two nodes, where each node has a long queue of frames to send with its Ethernet interface. Does your classmate Mac Phair s proposed change to the Ethernet MAC protocol eliminate the capture effect for such networks with more than two nodes? Why or why not? Again, in your answer, assume that the length of all frames transmitted is many Ethernet slot times. f. Upon hearing about Mac Phair s suggested change to the Ethernet MAC protocol, another of your 3035/GZ01 classmates, Sharon Doughty, worries that Ethernet senders using Phair s changed BEB will choose different mean backoffs after collisions than Ethernet senders using unmodified BEB. Consider a scenario where there are only two senders on an Ethernet, X and Y, both with long queues of frames to send. Consider the case where an unmodified sender X (without Phair s BEB change) encounters two successive collisions as it attempts to transmit a frame. Then consider the case where a modified sender (with Phair s BEB change) has transmitted one frame successfully, then encounters two successive collisions on the next frame it attempts to transmit. What are the expected mean backoff intervals (in slots) for the unmodified sender X and the modified sender X? In this scenario only (two senders only, each with infinitely many frames to send), will modified and unmodified Ethernet senders choose different mean backoffs after two collisions? Justify your answer by referring to the relevant details of stock Ethernet BEB and Phair s modified Ethernet BEB. [5 marks] [Question 1: Total 33 marks] COMPGZ01 3 TURN OVER

4 2. TCP a. The TCP sender retransmits data segments (packets) for which it does not receive ACKs. It also samples the elapsed round-trip time (RTT) between the sending of a data packet and the return of the ACK for that sequence number, as input to its estimate of the connection s RTT. Suppose the TCP sender retransmits a data packet, and subsequently receives an ACK for that packet. Why should the TCP sender not use RTT samples for retransmitted packets in its RTT estimate? b. Suppose you make repeated file transfers using TCP between two hosts on the Internet, A and B, where A acts as the sender and B as the receiver. You notice that the TCP throughput you are obtaining is quite poor, and decide to investigate. Your 3035/GZ01 classmate tells you that even if the Internet path(s) between A and B drops no packets, but merely reorders them, it s quite possible that you will observe very low-throughput TCP data transfers. Is your classmate correct? Justify your answer by either explaining precisely why TCP would suffer low throughput under packet reordering, or by explaining why it would not. Refer to the relevant specific details of the TCP protocol in your answer. [4 marks] c. Your 3035/GZ01 classmate tells you that you can determine whether the Internet path(s) between a TCP sender and receiver is (are) reordering data packets by detecting duplicate data packets (i.e., for the same TCP sequence number) that arrive at the receiver. He says that when significant reordering occurs, the receiver will receive duplicate data packets. Is your classmate correct? Justify your answer by referring to the relevant details of the TCP protocol. [4 marks] d. Suppose that a TCP sender has exited slow start and is in steady state, such that its congestion window consists of many dozens of packets, and is approximately equal to the bandwidth-delay product of the Internet path between itself and the TCP receiver. Suppose further that a single TCP ACK is dropped en route from the TCP receiver to the TCP sender. Your classmate predicts that the dropping of this TCP ACK will cause the TCP sender to experience a retransmit timeout (since the lost ACK won t arrive before the retransmit timer for the corresponding TCP data packet expires). Is your classmate correct? Justify your answer by referring to the relevant details of the TCP protocol. [4 marks] COMPGZ01 4 CONTINUED

5 e. Consider the below graph, which plots a TCP sender s congestion window size in segments (packets) vs. time measured in RTTs. Referring to the above graph, answer all the following questions: i. During which time periods is TCP slow start operating? [2 marks] ii. During which time periods is the TCP sender engaging in additive increase? [2 marks] COMPGZ01 5 TURN OVER

6 iii. Is the change in congestion window between RTT 14 and RTT 15 a fast retransmit or a timeout, what about the points at RTT 14 and RTT 15 lets you make that determination, and what observation at the sender triggers this change in congestion window? [3 marks] iv. Suppose the TCP sender experienced a retransmit timeout at RTT 18 (this is not shown in the figure; while answering this part of this question, disregard the points on the graph at RTTs 18 and greater). If a timeout occurred at RTT 18, how big would the TCP sender s congestion window need to grow before the TCP sender would exit slow start? Justify your answer by referring to the relevant details of the TCP congestion control algorithm. [2 marks] f. The system administrator at a large company wants to block malicious Internet traffic from reaching end hosts within the company s network. He adds a rule to the company s firewall (which sits at the one gateway link between the company s network and the rest of the Internet) that blocks all Internet Control Message Protocol (ICMP) packets arriving inbound to the company s network from the Internet, because ICMP packets are often malicious (e.g., sent by attackers attempting a denial of service (DOS) attack on the company s network). To his surprise, shortly after the system administrator adds this rule to the company s firewall, he receives a huge number of support requests from users of computers on the company s network. The users who send these support requests all say that when they attempt to send a large file to a destination elsewhere on the Internet with TCP, the connection hangs that is, the receiver elsewhere on the Internet stops receiving any data packets at all on the user s TCP connection, and the sender stops receiving any ACKs on the user s TCP connection. i. Why did the installation of this firewall rule cause this problem with outbound TCP data transfers? Justify your answer by referring specifically to the mechanism that this firewall rule breaks, and why the rule breaks that mechanism. [3 marks] ii. While leaving this same firewall rule in place (and not allowing any additional packets to traverse the firewall inbound), propose a change to the sender s TCP and IP protocol implementations that will fix the problem the users are reporting. Explain why this change fixes this problem, and comment on any undesirable side-effects that this fix may have. [3 marks] [Question 1: Total 33 marks] COMPGZ01 6 CONTINUED

7 3. BGP a. A tier-1 Internet service provider (ISP), A and its customer, a tier-2 ISP, B, connect at a site in London and a site in Germany. Each has an extensive international backbone. Say that UCL is ISP A s customer, and Volkswagen (in Germany) is ISP B s customer. If a host at UCL sends a packet to a host at Volkswagen, the packet must cross from A s backbone onto B s backbone at some connection point. i. What typical routing policy will determine where the packet crosses from ISP A s network onto ISP B s, where will this policy tend to make the crossing take place, and what is the economic motivation for this policy? [4 marks] ii. Suppose ISP B would like to force ISP A to carry packets bound for ISP B (and its customers) all the way to their connection point in Germany on ISP A s own infrastructure. What BGP mechanism would ISP B use to accomplish this goal, how would ISP B use it, and how would it affect BGP s behavior? [8 marks] b. When the network topology changes, does BGP ever find a route containing a loop among ASes? If so, give an example showing this behavior; if not, explain why it cannot. [3 marks] c. An AS may hear a prefix for the same destination announced in BGP via a provider, customer, and/or peer. What is the order of preference among these three, and why? d. In a typical tier-1 ISP, all routers run an interior gateway protocol (IGP), such as Link-State routing, and an exterior gateway protocol (EGP), such as ibgp. To forward packets to all of the Internet s destination prefixes, a router must have a forwarding table entry for every one of these prefixes. Describe how a router combines the information it learns from its IGP with that it learns from its EGP to produce this complete routing table, with one entry for every Internet destination prefix. e. Unlike distance-vector routing, BGP does not send periodic announcements of the contents of its routing table. Why did the designers of BGP make this choice, and how does BGP ensure that routers hold up-to-date routing tables without such periodic updates? Refer to the details of BGP s route propagation scheme in your answer. [Question 3: Total 33 marks] COMPGZ01 7 TURN OVER

8 Part TWO Multiple Choice Answer ALL questions in this section. The questions in this section are True/False. For each statement, mark the one appropriate box below ON YOUR EXAM PAPER. Using an HB pencil, fill in the rectangle on the Form A multiple choice answer sheet corresponding to whether each statement in each question is true ( T ) or false ( F ). Fill in D if you don t know. DO NOT GUESS. You will be awarded one mark for each statement correctly classified as true or false, but you will lose one mark for each choice incorrectly classified as true or false. You will neither gain nor lose marks if you fill in don t know. Statements for which you do not fill in an answer will be classified as don t know. Final results will be normalized across the class. 1. TCP Congestion Control Consider the Transmission Control Protocol (TCP) and its congestion control mechanisms. Which of the following statements are true? A. When two TCP flows compete at a single bottleneck and both experience the same packet loss rate, but have different round-trip times (RTTs), the flow with the shorter RTT will achieve a greater throughput. B. A TCP sender relies on ICMP packets from routers signaling in-network congestion to learn when to reduce its congestion window size. C. As the packet loss rate on a path increases, TCP s throughput decreases linearly in the packet loss rate. D. A retransmit timeout reduces TCP s congestion window by half. E. TCP s congestion control ensures that when a TCP flow and a non-tcp, non-congestioncontrolled flow compete at a bottleneck link, the TCP flow will always get a fair share of the bottleneck link s capacity. 2. Error Correction Which of the following are true statements about the various error detection and correction codes we have learned about? A. A 2 1 -rate repetition code can correct one bit error in a byte of data. B. A two-dimensional parity check can correct one bit error (assuming the receiver knows at most one bit is incorrect). C. A two-dimensional parity check can detect all three-bit errors (assuming the receiver knows at most three bits are incorrect). D. The Internet checksum can detect any two-bit error in a frame. E. With appropriate choice of polynomial, the cyclic redundancy check can detect any two-bit error in a frame. COMPGZ01 8 CONTINUED

9 3. MACAW/ MAC Which of the following statements about the MACAW and wireless MAC protocols are true? A. MACAW s request-to-send (RTS) and clear-to-send (CTS) mechanism mitigates the exposed terminal problem. B. The MAC protocol s inclusion of link-layer reliability is a flagrant violation of the end-to-end argument. C. The MACAW and MAC protocols require all nodes to listen with their radios at all times (except while they are sending). D. Carrier sense causes the hidden terminal problem (i.e., the hidden terminal problem would not exist if MACAW did not use carrier sense). E. Carrier sense causes the exposed terminal problem (i.e., the exposed terminal problem would not exist if MACAW did not use carrier sense). 4. BGP Which of the following are true statements about the Border Gateway Protocol (BGP)? A. A BGP router may peer with another BGP router that is more than one physical hop away. B. A network running BGP that hears a route from a customer AS will normally forward it on to a neighbouring provider AS. C. A network running BGP that hears a route from a provider AS will normally forward it on to another different neighbouring provider AS. D. A network running BGP that hears a route from a peer AS will normally forward it on to another different neighbouring peer AS. E. A BGP router will that hears two copies of a route from two different neighbouring ASes will normally prefer the route from a customer AS over that from a provider AS. COMPGZ01 9 TURN OVER

10 5. TCP Protocol and Congestion Control A. TCP practices self-clocking transmission by using a timer at the sender to pace packet transmissions at a perfectly smooth rate. B. Over a network that can drop packets, the TCP protocol ensures using FIN messages that once both endpoints of a TCP connection decide to end the connection, they will always both reach agreement that both have reached this decision. C. For very short transfers of fewer than 4 packets, TCP cannot invoke fast retransmit to recover from losses. D. In normal operation, a TCP sender repeatedly provokes loss of its packets at one or more queues along the path between sender and receiver. E. TCP begins new connections with an initial sequence number of zero. 6. Spanning Tree Protocol Which of the following are true statements about the Ethernet Spanning Tree Protocol (STP)? A. All switches in an extended Ethernet that is running STP will eventually agree on a single root switch. B. When the STP converges, all links out of the root switch will be in the root state. C. When the STP converges, no switch will have all its ports in the designated state. D. Switches compare configuration messages first based on lower distance to root, then based on lower root identifier. E. Link failures can briefly introduce forwarding loops until configuration messages time-to-live fields expire. COMPGZ01 10 CONTINUED

11 7. BGP Routing Which of the following statements about the Border Gateway Protocol (BGP) are correct? A. BGP may choose routes that contain loops immediately after a topology change. B. Given three autonomous systems (ASes), A, B, and C, consider the round-trip time (RTT) of the path selected by BGP between A and B with the sum of the round-trip times of the paths selected by BGP between A and C and between C and B. BGP never selects paths such that the RTT of the former path (A B) is greater than the sum of the RTTs of the latter paths (A B and B C). C. If one tier-1 ISP stops peering with another tier-1 ISP, no customers of either ISP will become partitioned from one another, because BGP will then begin to route traffic between these two tier-1 ISPs via some third tier-1 ISP. D. BGP routers in an AS must honor multi-exit discriminator (MED) attributes advertised by routers in neighboring ASes. E. Under BGP, routers not on an AS s border (i.e., those connected only to other routers within the same AS, with no links to routers in other ASes) need not store routing tables with entries for all destination prefixes in the global Internet MAC Consider the IEEE medium access control (MAC) protocol. Which of the following statements are true? A. Request to Send (RTS) and Clear to Send (CTS) packets are of no use in mitigating the hidden terminal problem when senders attempt to send very short data packets. B senders use collision detection to share the wireless medium more efficiently. C. Suppose two senders, A and C, wish to send concurrently to two receivers, B and D, respectively. Carrier sense may reduce the aggregate throughput achieved by A and C in some topologies, as compared with the aggregate throughput achieved when neither uses carrier sense. D. Address Resolution Protocol (ARP) request packets sent on networks are always acknowledged at the link layer. E. An sender backs off exponentially before retransmitting when it does not receive a link-layer acknowledgement (ACK) for its unicast data transmission. [Total 40 marks, after normalization of marks] END OF PAPER COMPGZ01 11