Deck 15: Understanding Multiprocessor Organizations and System Performance

Question 1

__________ is an example of a responsibility for an operating system in a multiprocessing system that does not exist in a uniprocessor system.

A) Ensuring that processes are evenly distributed on the system's processors
B) Ensuring that processors operate on consistent copies of data stored throughout the system
C) Synchronizing execution of related processes
D) all of the above

Answer 1

D

Question 2

A __________ is an example of a SIMD processor?

A) Vector processor
B) Dual-processor Intel Pentium system
C) Uniprocessor with a pipeline, containing many stages
D) all of the above

Answer 2

A

Question 3

__________ is not an example of a technique for adding parallelism into a SISD computer?

A) Pipelining
B) Superscalar execution
C) VLIW execution
D) none of the above

Answer 3

D

Question 4

The __________ is a measure of the performance of an interconnection scheme.

A) bisection width
B) network diameter
C) node degree
D) number of nodes in a system

Question 5

For a system with 16 nodes, which of the following interconnection schemes would have the smallest network diameter?

A) 4-connected 2-D mesh network
B) hypercube
C) crossbar-switch matrix
D) multistage network

Question 6

Which of the following statements is a characteristic of a tightly coupled system?

A) Components are connected indirectly through communication links.
B) Swapping components in to and out of the system is easier than on a loosely coupled system.
C) Contention can build up quickly at shared resources.
D) Communication is generally accomplished through message passing or shared virtual memory, but not shared physical memory.

Question 7

A loosely coupled system is __________ than a tightly coupled system.

A) less fault tolerant
B) more flexible
C) more efficient
D) more expensive

Question 8

In which type of environment is a master/slave multiprocessor organization most appropriate?

A) scientific computing applications in which large matrices are manipulated
B) cluster of workstations
C) interactive environment
D) mission-critical environment

Question 9

Master/slave multiprocessor organizations __________.

A) exhibit high fault tolerance
B) enable many processors to execute the operating system, reducing the likelihood that contention will build up for operating system resources.
C) exploit hardware asymmetry (e.g., if one processor is more powerful than other processors) and known parallelism in computationally-intensive tasks.
D) exhibit high scalability, where adding a processor increases the system throughput by the new processor's rated capacity.

Question 10

A separate kernels multiprocessor organization does not __________.

A) prevent information from being shared among the various operating systems
B) exhibit high system fault tolerance
C) permit only minimal cooperation between processors
D) limit resource sharing

Question 11

A separate kernels multiprocessor organization is more appropriate for __________.

A) a dual-processor personal computer
B) scientific computing applications in which large matrices are manipulated
C) streaming video processing
D) a transaction processing system that requires high fault tolerance, but little coordination

Question 12

__________ is the most fault-tolerant type of multiprocessor organization.

A) Master/slave
B) Separate kernels
C) Symmetrical
D) All of the above exhibit similar levels of fault tolerance.

Question 13

A symmetrical multiprocessor organization must __________, unlike master/slave and separate kernels organizations.

A) ensure mutual exclusion to all kernel data
B) be able to recover from a processor failure
C) handle all I/O requests using only one processor
D) all of the above

Question 14

In a UMA multiprocessor, __________.

A) shared virtual memory presents the illusion of shared physical memory, although physical memory is not shared.
B) memory access time for a given processor varies depending on which memory module contains the requested data item.
C) memory is viewed as a cache so that data can be migrated at the granularity of a memory line, causing a lower percentage of cache misses to be serviced remotely.
D) memory access time is nearly the same for all requested data items in main memory.

Question 15

In an SMP system, __________ is shared between all processors.

A) memory
B) I/O devices
C) processes
D) all of the above

Question 16

A NUMA system more suitable for __________.

A) a dual-processor personal computer
B) a cluster of workstations
C) a system with a large number of processors (greater than 64) inside a single machine
D) a large, distributed computer such as Japan's Earth Simulator

Question 17

Why is the NUMA design more scalable than the UMA design?

A) NUMA systems are loosely coupled, whereas UMA systems are tightly coupled.
B) NUMA systems reduce bus collisions when each processor's local memory services most of the processor's memory requests. By contrast, UMA systems can saturate the shared bus quickly.
C) NUMA systems use more than one memory module, whereas UMA systems use only one.
D) Memory access time is faster in NUMA systems than in UMA systems.

Question 18

Which of the following statements does not describe a benefit of the COMA design?

A) COMA systems require a less complicated memory coherency protocol than NUMA systems.
B) False sharing is reduced compared to NUMA systems that use page migration.
C) The number of cache misses serviced remotely is reduced compared to NUMA systems.
D) Operating system designers do not have to implement page migration and replication strategies.

Question 19

Attraction memory is __________.

A) memory shared equally between all processors
B) serves as a shared L3 cache for processors in a multiprocessor system
C) memory, associated with COMA nodes, that is organized as a cache
D) a memory system that uses page migration and/or page replication

Question 20

Nodes in a NORMA multiprocessor communicate through __________.

A) explicit messages and, in some systems, shared virtual memory (SVM)
B) explicit messages and, in some systems, shared physical memory
C) SVM and, in some systems, shared physical memory
D) explicit messages, SVM and, in some systems, shared physical memory

Question 21

A __________ is an example of a NORMA system.

A) network of computers
B) system with a large number of processors that share physical memory. Access to memory modules local to a processor is much faster than access to global memory modules.
C) distributed system
D) dual-processor personal computer

Question 22

A system exhibits __________ if the value obtained from reading a memory address is always the value that was most recently written to that address

A) data reliability
B) memory coherence
C) data integrity
D) cache consistency

Question 23

How is bus snooping implemented to ensure cache coherence?

A) The system maintains a centralized directory that records the items in each processor's cache. On a cache write, the writing processor "snoops" this list to determine whether the item is in other processors' caches, and if so, tells these other processors to remove the item.
B) The system only allows one processor at a time to cache each data item; when a processor tries to cache a data item, it "snoops" other processors' caches over the bus to determine whether it is allowed to cache the data.
C) Each processor "snoops" the processor-memory bus to determine whether another processor has sent a message about an update to a data item stored in its cache.
D) Each processor "snoops" the bus to determine whether a requested write from another processor is for a data item in the processor's cache. If the data resides in the processor's cache, the processor removes the data item from its cache.

Question 24

How is the CC-NUMA home-based cache coherency protocol implemented when a cache miss occurs on a read?

A) The reading node contacts the home node, which always stores the most updated version of the data item. The home node forwards the data to the requesting node.
B) The reading node contacts the home node. If the data item is clean, the home node forwards it to the reading node. If the data item is dirty, the home node forwards the request to the node with the dirty copy. This node sends the data to both the home node and the requesting node.
C) The reading node contacts the home node. If the data item is clean, the home node forwards it to the reading node. If the data item is dirty, the home node broadcasts a request for the dirty copy to all nodes. The node with the dirty copy sends the data to both the home node and the requesting node.
D) The reading node contacts the home node. If the data item is clean, the home node forwards it to the reading node. If the data item is dirty, the home node broadcasts a request for the dirty copy to all nodes. The node with the dirty copy sends the data to the home node, which forwards it to the requesting node.

Question 25

A page that is __________ is a good candidate for page migration.

A) frequently read by several remote nodes, but not modified
B) frequently modified by several remote nodes
C) frequently modified by a single remote node
D) frequently read and modified by several remote nodes

Question 26

How might page migration and page replication actually reduce system performance?

A) Page replication requires a more complex coherency protocol because the same page can exist at separate nodes.
B) Replicating or migrating a page is more costly than remotely referencing a page a small number of times. If pages that are replicated or migrated are not referenced again at their new nodes, then these strategies slow average memory access time.
C) The information that must be maintained to make good migration and replication decisions can consume substantial memory overhead.
D) all of the above

Question 27

In the __________ memory coherence protocol, a write causes all nodes except the writing node to remove their copy of the modified page.

A) invalidation
B) write broadcast
C) bus snooping
D) strict consistency

Question 28

In the __________ implementation of a memory coherence protocol, all shared page updates that occur between an acquire operation and a release operation are batched as one update message after the release operation completes.

A) lazy release consistency
B) delayed consistency
C) home-based consistency
D) release consistency

Question 29

Processor affinity is the __________.

A) relationship among processes that execute on the same processor
B) relationship of a process to a particular processor and its local memory and cache
C) likelihood that a process will execute on a particular processor
D) likelihood that a memory page will reside in the local memory of a particular processor

Question 30

__________ process scheduling attempts to exploit parallelism inherent in groups of related processes.

A) Job-blind scheduling
B) Space-partitioning scheduling
C) Timesharing scheduling
D) RRprocess scheduling

Question 31

Which of the following job-blind process scheduling algorithms attempt to exploit processor affinity?

A) FIFO scheduling
B) RRprocess scheduling
C) SPF scheduling
D) none of the above

Question 32

Which job-blind process scheduling algorithm is least likely to indefinitely postpone a process?

A) FIFO scheduling
B) RRprocess scheduling
C) SPF scheduling
D) The answer depends on the length of the runtime queue and the characteristics of each process so cannot be determined.

Question 33

Consider a system containing 18 processors and three jobs: job A contains 11 runnable processes, job B contains eight runnable processes and job C contains four runnable processes. The system contains 18 processors. If the system uses dynamic partitioning scheduling, how are processors allocated to jobs?

A) A receives seven processors; B receives seven processors; and C receives four processors.
B) A receives eight processors; B receives six processors; and C receives four processors.
C) Each job receives six processors.
D) Each job receives four processors; the remaining six processors are reserved for incoming jobs.

Question 34

The undivided coscheduling algorithm does not __________.

A) permit indefinite postponement
B) exploit parallelism between related processes
C) exploit processor affinity
D) permit processors to remain idle if there are processes waiting to execute.

Question 35

A system might decide to migrate a process __________.

A) to balance processor load
B) to reduce the cost of communication between processes that execute at different nodes before migration
C) to promote resource sharing
D) all of the above

Question 36

__________ are not migrated with a process as part of its process state.

A) Machine registers contents
B) Children processes
C) The state of the process's open files
D) Working set memory pages

Question 37

How does a system typically handle messages sent to a migrating process?

A) All messages must be sent to the receiving node once migration starts; all messages sent to the sending node are lost.
B) The sending node creates a dummy process that receives messages for the migrating process. Once migration completes, the sending and receiving nodes inform other nodes of the process's new location. The dummy process forwards messages to the migrated process, and the sending node terminates the dummy process.
C) All messages are sent to the sending node, which forwards the messages to the new node and informs the message's sender of the process's new location.
D) The sending node creates a message queue to hold all messages destined for the migrating process. Once migration completes, the sending and receiving nodes inform other nodes of the process's new location. The sending node forwards the messages in the message queue to the receiving node and destroys its instance of the process.

Question 38

A migrated process's reliance on its previous node is called __________.

A) residual dependency
B) lazy dependency
C) migration dependency
D) temporal dependency

Question 39

__________ is not a desirable characteristic of a successful migration strategy.

A) Transparency
B) Scalability
C) Residual dependency
D) Fault tolerance

Question 40

In __________, the sender begins transferring dirty pages before the original process is suspended, then migrates the process when the number of dirty page remaining at the sending node reaches a lower threshold.

A) dirty eager migration
B) precopy migration
C) lazy migration
D) flushing migration

Question 41

An advantage of eager migration over dirty eager migration is __________.

A) less initial migration latency
B) that the migrating process will have more of its working set in memory immediately after it migrates
C) that no messages are lost during migration
D) none of the above

Question 42

In which type of environment would static load balancing be most appropriate?

A) interactive timesharing system
B) scientific computing
C) UMA multiprocessor
D) Web server

Question 43

__________ is a goal of static load balancing algorithms.

A) Ensuring that no processor remains idle while there is a process waiting to execute
B) Executing processes that communicate at the same node
C) Reducing the variance of response times
D) all of the above

Question 44

In which environments does a symmetric load balancing policy perform well?

A) The system is always overloaded.
B) The system is always underloaded.
C) There are some periods when the system is underloaded, and some when the system is overloaded.
D) all of the above

Question 45

In the __________ load balancing algorithm, processors classify their loads as low, normal or high, send this information to other processors and use a receiver-initiated policy based on these classifications to migrate processes.

A) drafting
B) bidding
C) diffusion
D) symmetric

Question 46

A process waiting for a spin lock __________.

A) blocks
B) busy-waits for the spin lock.
C) busy-waits for a finite amount of time, then it blocks
D) preempts the lock holder if its priority is higher than the lock holder's priority

Question 47

An adaptive lock is a mutual exclusion lock that __________.

A) can either be held by one process in exclusive mode or several processes in shared mode
B) allows a process to specify the amount of time it will hold the lock; other waiting processes can use this time to determine whether to spin or block.
C) allows the system to change whether the lock is spinning or blocking based on the system load
D) allows the system to switch between allowing only kernel-mode components to hold the lock and allowing all processes to hold the lock

Question 48

A __________ occurs when two or more processes access a resource in a nondeterministic order, and the outcome depends on the order that the processes access the resource.

A) thundering herd
B) mutual exclusion violation
C) residual dependency
D) race condition

Question 49

A(n) __________ lock is useful for a situation in which__________

A) spin lock; the system load is high
B) read/write lock; the protected resource is often accessed without being altered
C) sleep/wakeup lock; system load is light.
D) adaptive lock; system load is static.