Quiz #4

Question 1

What is the relationship between throughput and latency?

Answer 1

1. Latency is the time it takes for a packet to go from point A to point B, while throughput is the number of packets per unit time that can be sent between point A and point B
2. Lower latency implies
   greater throughput, but no the other way around.

Question 2

What is utilization?

Answer 2

Utilization has to do with how well the resources in the system as a whole are being utilized.
How well the networking infrastructure is used.

Question 3

What is Scalability in the context of DC networks?

Answer 3

As you increase the offered load into the system, is the performance that is experienced by the applications remaining unchanged?

Question 4

In regard to performance evaluation, what is the difference between Modeling and Simulation?

Answer 4

Modeling is mathematically expressing the behavior of the system while simulation is constructing a computer program that mimics the behavior of the system.

Modeling is a quicker way of validating design ideas than simulation

Question 5

What can go wrong in DC networks?

Answer 5

1. Forwarding loops
2. Link failures
3. Forwarding inconsistencies - often leads to forwarding loops.
4. Unreachable hosts (network could get fragmented due to failed links and this may lead to a partition in the network)
5. Unwarranted access to hosts

Question 6

What is forwarding loops?

Answer 6

you expect that every hop of the packet is making an advanced towards the destination. But if that is not happening and it keeps cycling back and forth, that’s the forwarding loop. That could be because of the tables being incorrectly set.

Question 7

What is the reason for forwarding inconsistencies?

Answer 7

this has to do with the fact that in a large network, all the tables are not getting updated at the same time, it is happening at different times. There could be inconsistencies between the forwarding tables from one switch to the next twitch.

Question 8

How domain-specific languages can be useful to testing and debugging?

Answer 8

you want domain-specific languages that limit the kinds of errors that you can potentially get in the way the control is setup. It also means that you have a limited set of primitives that are used for programming these switches.

Question 9

What is the limitation of model checking?

Answer 9

Model checking (in its pristine form), is unscalable to large systems because there are too many state pieces.

*You have to constrain the state spaces so that you can actually test and debug your system, and that’s why combining model checking technique with symbolic execution is an appealing way of dealing with testing and debugging of data center networks.

Question 10

What are the effects of network bugs?

Answer 10

1. Unauthorized entry of packets into a secured zone
2. Vulnerability of services and the infrastructure to attacks
3. Denial of critical services
4. Affect network performance and violation of SLAs (system level agreements)

Question 11

What are the elements of NICE’s secret sauce?

Answer 11

1. State-Space exploration via Model Checking (MC)

2. Combine Symbolic Execution (SE) with Model Checking to prevent state-space explosion.

Question 12

What are the keys to scalability of OFRewind?

Answer 12

*OFRewind is a static analysis tool of OpenFlow programs.

1. Record only control plane traffic
2. Skip/aggregate data plan traffic
3. Best effort replay as opposed to deterministic (because the idea is to identify whether there was any performance problems/correctness issues in the behavior).

Question 13

What constitutes a network breakpoint in NDB? Select

all that apply

Answer 13

NDB = Network Debugger.
Breakpoint is a filter on packet header, e.g.
Switches send “postcard” on matching entries to a central collector.
Collector stores the postcards to construct a “backtrace”.

Question 14

What are the main elements of Header Space

Analysis? Select all that apply

Answer 14

1. Packet (with a header size of L) as a point in a L-dimensional hyperspace
2. Switches as transformers causing state transition of the packet headers - T3(T2(T1(h)))

Question 15

What are the common sources of “small” packets in the DC network?

Answer 15

*Packets with length < 200 bytes are considered small, and from all packets, there ~50% such packets.

The common sources are packets like TCP acks and “I am alive” msg, i.e. control messages.

Question 16

Most traffic in DC networks tend to be within a rack. Why?

Answer 16

75% of traffic is within a rack - because of good colocation of application components.

Question 17

What are the indicators that there is sufficient Bisection bandwidth in DC networks?

Answer 17

1. Only small fraction of the available bisection used (<30%).
2. Small packet sizes dominate

Question 18

What are the indicators that links are under-subscribed?

Answer 18

1. 75% traffic within a rack.

2. core links utilization < 25%.