Name of College/ University NIET Greater Noida

Course Code RCS-083

Exam Name Parallel and Distributed computing
Total Question Pool 175
Question Difficulty Category, Questions Marks
Type Subcategory
Single Easy Parallel 1.The number of tasks into which a problem is 1
choice Algorithms, decomposed
Decomposition Determines its
A. Functionality
UNIT-I B. Completeness
C. Balancing
D. Granularity

Single Medium Parallel 2. The longest path in a task dependency graph is called 1
choice Algorithms, the
Task A. Directed Path
Interconnection B. Sequence of Task
C. Critical Path Length
D. Critical Path

Single Medium Parallel 3. Appropriate mapping of tasks to processes is critical to 1

choice Algorithms, the parallel performance of an algorithm. These Mappings
Mapping are determined by
A. task dependency graph
B. task interaction graph
C. A & B both
D. Process graph

Single Easy Parallel 4. A set of commonly used Decomposition techniques 1

choice Algorithms, that apply to broad classes of problems are:
Mapping A. recursive decomposition
Technique B. data decomposition
C. exploratory decomposition
D. All of above
Question Difficult Category, Questions Marks
Type y Subcategory
Single Easy Parallel 5. Recursive Decomposition suited to problems that are 1
choice Algorithm, solved using the
Decomposition A. Branch and Bound
B. Divide and Conquer
C. Backtracking
D. Iterative Deepening
Single Medium Parallel 6. Computation can be viewed as a sequence of 1
choice Algorithm, transformation from the input to the output data. In these
Decomposition cases, it is often beneficial to use one of the intermediate
 stages as a basis for decomposition.
A. True
B. false

Single Medium Parallel 7. The Owner Computes Rule generally states that the 1
choice Algorithm, process assigned a particular data item is responsible for
Decomposition A. all computation
B. computations that use the input data
C. computations that use the Output data
D. None
Multiple Easy Parallel 8. Once a problem has been decomposed into independent 1
choice Algorithm, tasks, the characteristics of these tasks critically impact
Task choice and performance of parallel algorithms. Relevant task
characterizatio characteristics include:
n A. Task generation
B. Task Size
C. Size of Data associated with Task
D. None

Single Easy Parallel 9. Once a problem has been decomposed into concurrent 1
choice Algorithm, tasks, these must be mapped to processes. Then mapping
Decomposition must minimize
A. Task interaction
B. Overheads
C. A & B both
D. None
Single Medium Parallel 10. In Mapping techniques for minimum idling the task are 1
choice Algorithm, mapped to a process a-priori in
Task Mapping A. Dynamic Mapping
B. Static mapping
C. Regular Mapping
D. Irregular Mapping
Single Easy Parallel 11. Instructions in a program may be related to each other. 1
choice programming, The results of an instruction may be required for subsequent
Pipelining instructions. This is referred to as
A. True Dependency
B. Resource Dependency
C. Instruction Dependency
D. Procedure Dependency
Single Easy Parallel 12. The rate at which data can be pumped from the memory 1
choice Programming, to the processor determines the
Memory A. Latency
B. Bandwidth
C. Transfer rate
D. None
Single Easy Parallel 13. The improvement in performance resulting from the 1
choice Programming, presence of the cache is based on the assumption that there is
Memory repeated reference to the same data item. This notion of
repeated
reference to a data item in a small time window is called
 A. Spatial Locality
B. Reference locality
C. Temporal Locality
D. None
Single Difficult Parallel 14. A computer has a single cache (off-chip) with a 2 ns hit 2
choice Programming, time and a 98% hit rate. Main memory has a 40 ns access
Memory time. If we add an on-chip cache with a .5 ns hit time and a
94% hit rate, what is the computer’s effective access time?
How much of a speedup does the on-chip cache give the
computer?
A. 2.4 ns, 4.2 ns
B. 2.6 ns,4.6 ns
C. 2.8 ns,4.2 ns
D. None
Single Easy Parallel 15. When the data at a location in cache is different from the 1
choice Programming, data located in the main memory, the cache is called
Memory A. Unique
B. Inconsistent
C. Fault
D. Variable
Single Medium Parallel 16. Imagine sitting at your computer browsing the web 1
choice Programming, during peak network traffic hours. We access a whole bunch
Network of pages in one go – amortizing the latency across various
accesses. This approach is called
A. Multithreading
B. Prefetching
C. Spatial Locality
D. Temporal Locality

Single Easy Parallel 17. Computers in which each processing element is capable 1
choice Programming, of executing a different program independent of the other
Architecture processing elements are called
A. SIMD
B. MISD
C. SISD
D. MIMD

Single Medium Parallel 18. If the time taken to access certain memory words is 1
choice Programming, longer than others, the platform is called
Architecture A. UMA
B. NUMA
C. SPMD
D. CUDA
Single Medium Parallel 19. Dynamic networks for parallel Computers , are built 1
choice Programming, using
Network A. Point to Point Link
B. Switches and Communication links
C. A & B both
 D. None of the above
Single Easy Parallel 20. The minimum volume of communication allowed 1
choice Programming, between any two halves of the network is called
Network A. Section bandwidth
B. Cross Section bandwidth
C. Bisection bandwidth
D. None

UNIT –II

Questio Difficulty Category, Questions Marks

n Type Subcategory
Single Easy CUDA, 21. CUDA allows threads in the same block to coordinate 1
choice Transparent their activities using a barrier synchronization function
Scalability A. SynchThread()
B. Synch()
C. SyncThread()
D. SyncThreads()
Single Easy CUDA, Thread 22. Threads being blocked altogether and being executed in 1
choice the sets of 32 Threads called a
A. Thread Block
B. 32 Thread
C. 32 Block
D. Unit Block
Single Medium CUDA, Thread 23. In CUDA Thread Organization blockId and threadId 1
choice Organization are:
A. User Defined Variables
B. Derived Variables
C. Built in Variables
D. Global Variables
Single Medium CUDA, 24. CUDA provides two memory management functions. 1
choice Memory Those functions are:
Management A. CudaAlloc(), CudaFree()
B. CudaMalloc(), CudaFree()
C. A & B both
D. None of the above
Single Medium CUDA, API 25. Can GPUs run hundreds or thousands of threads in 1
choice parallel and has its own DRAM?
A. True
B. False
Single Easy CUDA, API 26. A new hardware and software architecture for issuing 1
choice and managing computations on the GPU is called:
A. Parallel Processor
B. GPU Processor
C. Compute Unified Device Architecture
D. None of above
Single Medium CUDA, Kernel 27. Triple angle bracket in mykernel<<1,1>>(); function in 1
choice CUDA mark a call from
A. Host code to device code
 B. Device code to Host code
C. Server to client code
D. Client to server code
Single Difficult CUDA, Kernel 28. add<<< 256, 1>>>() function in Kernel computing 1
choice A. Create 256 copies of the kernel and runs them in
parallel
B. Create 256 copies of the kernel and run then
sequentially
C. Only creates 256 copies of kernel
D. None of above
Single Medium CUDA, 29. The number of threads per block with which we can 1
choice Combining launch a kernel are:
Block and A. 65535
Threads B. 512
C. 2056
D. 512 or 1024

Single Easy CUDA, 30. Blocking Optimization is used to improve temporal 1

choice Memory locality , for reduce
A. Hit misses
B. Misses
C. Hit rate
D. Cache Misses
Single Easy CUDA, 31. The aggregate register file has a size of 1
choice Memory A. 1 MB
B. 2 MB
C. 3 MB
D. 4 MB
Single Medium CUDA, Thread 32. A code, known as Grid, which runs on a GPU consisting 1
choice of a set of
A. 32 Thread
B. Unit Block
C. 32 Block
D. Thread Block
Multiple Difficult CUDA, 33. CUDA makes following pieces of information available 2
choice Programming to each thread
Model A. The thread index (threadIdx)
B. The block index (blockIdx)
C. The size and shape of a block (blockDim)
D. The size and shape of a grid (gridDim)
Single Easy CUDA, GPU 34. The CPU and GPU have separate memory spaces. This 1
choice means that data that is processed by the GPU must be
moved from the CPU to the GPU before the computation
starts, and the results of the computation must be moved
back to the CPU once processing has completed.
A. True
B. False
Single Medium CUDA, 35. The CUDA parallel programming model has following 1
choice Parallel key abstractions at its core:
 algorithm A. a hierarchy of thread groups
B. shared memories
C. barrier synchronization
D. All of above
Single Easy CUDA, 36. The CUDA abstractions provides: 1
choice Features A. fine-grained data parallelism 
B.  thread parallelism
C. A & B both
D. None of above
Single Easy CUDA, Kernel 37. A kernel is executed in parallel by an array of threads 1
choice and all threads run the different code.
A. True
B. False
Single Medium CUDA, Grids 38. Threads are arranged as a grid of thread blocks and 1
choice different kernels can have
A. Different grid/block configuration
B. Same grid/block configuration
C. Both of above
D. None
Single Easy CUDA, Block 39.  Threads within a block can cooperate by sharing data 1
choice through some shared memory and by synchronizing their
execution to coordinate memory accesses.
A. True
B. False
Single Medium CUDA, Grid, 40. The grid of blocks and the thread blocks can be 1
choice Block A. 1 D
B. 2 D
C. 3 D
D. All of above

UNIT-III

Question Difficult Category, Questions Marks

Type y Subcategory
Single Easy Analytical 41. Overhead in a parallel system means 1
choice Modeling of A. processing in parallel systems
Parallel B. inter process interaction is less
algorithm, C. parallel scheduling is required
Source of D. elevated processing in parallel systems
overhead
Single Easy Performance 42. The time elapsed between the beginning and the 1
choice Metric, end of its
Execution execution on a sequential computer
runtime A. Parallel run time
B. Serial runtime
C. Communication overhead
D. None
Single Easy Performance 43. Overhead function is given by 1
choice Metric, Total A. t(o) = t(s) + t(o) - pt(p)
parallel B. t(o) = pt(p) - t(s)
overhead C. t(o) = t(s) - pt(p)
D. t(o) = pt(p) + t(s)
Single Medium Performance 44. The ratio of the time taken to solve a problem on 1
choice metric, a single processing element to the time
speedup required to solve the same problem on a parallel
computer with p identical processing elements is
called
A. Speed up
B. Overhead
C. Both of above
D. None
Single Difficult Performance 45. A problem can be solved in O(n) time on a single 1
choice metrics, processing element, its speedup is
Speedup A. Ɵ(n/log n)
B. Ɵ(n2/log n)
C. Ɵ(n3/log log n)
D. None
Single Easy Asymptotic 46. In parallel systems which are in practice, speed 1
choice analysis of up is ______ and efficiency is between ______
parallel A. greater than n, 1 and 2
program, B. equal to n , 0 and 1
speed up C. less than n, 0 and 1
D. less than n, 1 and 2
Single Medium Scalability 47. __________metric is defined as the speedup 1
choice metric, scaled obtained when the problem size is increased linearly
speedup with the number of processing elements.
A. Linear speedup
B. Scaled speedup
C. Non linear speedup
D. None of the above
Single Medium Scalability of 48. If the degree of concurrency of an algorithm is 1
choice parallel less than Q(W), then the isoefficiency function due to
system, concurrency is worse than Q(p).
isoefficiency A. True
B. False
Single Difficult Scalability of 49. If this serial component of the program takes time 1
choice parallel tserial, then during this time all the other processing
system, elements must be idle. This corresponds to a total
scaling overhead function of
characteristic A. (p - 1) x tserial
 B. (np-1)x tserial
C. (p-2) x tserial
E. None
Single Easy Scalability of 50. The scalability of a parallel system is a measure 1
choice parallel of its
system, capacity to decrease speedup in proportion to the
scaling number of processing elements
characteristic A. True
B. False

Single Easy Performance 51. Is it necessary for a parallel system to create 1

choice metric, overheads
overhead A. True
B. False
Single Difficult Speedup, 52. Consider the execution of a parallel program on a 2
choice super linearity two-processor parallel system. The
effect on program attempts to solve a problem instance of size
cache W. With this size and available cache of 64 KB on
one processor, the program has a cache hit rate of
80%. Assuming the latency to cache of 2 ns and
latency to DRAM of 100 ns, the effective memory
access time is
A. 2.26ns
B. 2.16ns
C. 2.36 ns
D. None
Single Easy Performance 53. Presence of a serial component in a system is 1
choice metric, which type of overhead
overhead A. interprocess interaction
B. excess computation
C. synchronization
D. idling
Single Medium Speedup, 54. Which isoefficiency function is best for the 1
choice super linearity parallel systems?
A. Large isoefficiency function
B. Moderate isoefficiency function
C. Small isoefficiency function
D. very small isoefficiency function
Single Easy Performance 55. Parallel execution time is expressed in terms of 1
choice Metric, A. cost, overhead and processing elements
Execution B. execution time and serial time
runtime C. efficiency and cost
D. processing elements, overhead and problem
size
Single Easy Performance 56. In parallel systems, to solve a problem, we use 1
choice Metric, A. 1 processing element
Execution B. n number of processing elements
runtime C.1 serial element and n number of processing
elements
D. only serial elements
Single Medium Performance 57. When p is less than n in a parallel system, is the 1
choice Metric, Cost system cost optimal?
A. True
B.False
Single Medium Performance 58. A cost optimal parallel system has an efficiency 1
choice Metric, Cost of
A. big oh of (0)
B. theta of (0)
C. omega of (1)
D. theta of (1)
Single Medium 59. A serial run time of a parallel system is 1
choice A. time elapsed between the starting and ending
of a process execution
B. time elapsed between processing elements
execution
C. total time required by a system to execute a
process
D. time elapsed between starting and ending of the
parallel process
Single Medium Analytical 60. Analytical modelling of a program refers 1
choice modeling A. to have a solution of analyzing the program
B. to have a solution of multiple parts of the
program
C. to have a solution based on the problem and
the solution describes the changes in the system
D. to have a solution based on the parallel
algorithms

UNIT IV

Question Difficulty Category, Questions Marks

Type Subcategory
Single Easy Dense 61. Which matrix has most of the elements (not all) as 1
choice matrix Zero?
algorithm, A. Identity Matrix
matrix B. Unit Matrix
C. Sparse Matrix
D. Zero Matrix
Single Medium Dense 62.  What is the relation between Sparsity and Density 1
choice matrix of a matrix?
algorithm, A. Sparsity = 1 – Density
matrix B. Sparsity = 1 + Density
C. Sparsity = Density*Total number of elements
D.Sparsity = Density/Total number of elements
Single medium Dense  63. Is O(n) the Worst case Time Complexity for 1
choice matrix addition of two Sparse Matrix?
algorithm, A. True
matrix B. False

Single Medium Dense 64. The matrix contains m rows and n columns. The 1
choice matrix matrix is called Sparse Matrix if ________
algorithm, A. Total number of Zero elements > (m*n)/2
matrix B. Total number of Zero elements = m + n
C. Total number of Zero elements = m/n
D. Total number of Zero elements = m-n
Single Medium Dense 65. Which of the following is not the method to 1
choice matrix represent Sparse Matrix?
algorithm, A. Dictionary of Keys
matrix B. Linked List
C. Array
D. Heap
Single Easy Dense 66. Is Sparse Matrix also known as Dense Matrix? 1
choice matrix
algorithm, A. True
matrix B. False
Single Easy Dense 67. Which one of the following is a Special Sparse 1
choice matrix Matrix?
algorithm, A. Band Matrix
matrix B. Skew Matrix
C. Null matrix
D. Unit matrix
Single Medium Dense 68. In what way the Symmetry Sparse Matrix can be 1
choice matrix stored efficiently?
algorithm, A. Heap
matrix B. Binary tree
C. Hash table
D. Adjacency List
Single Difficult Dense 69. What does the following piece of code do? 1
choice matrix
algorithm, for(int i = 0; i < row; i++)
array {
operation for(int j = 0; j < column; j++)
{
if(i == j)
sum = sum + (array[i][j]);
}
}
 System.out.println(sum);
A. Normal of a matrix
B. Trace of a matrix
C. Square of a matrix
D. Transpose of a matrix

Single Difficult Graph 70. In the given graph identify the cut vertices. 1
choice algorithm,
features

A. B and E
B. C and D
C. A and E
D. C and B
Single Difficult Graph 71. For the given graph(G), which of the following 1
choice algorithm, statements is true?
features

A. G is a complete graph
B. G is not a connected graph
C. The vertex connectivity of the graph is 2
D. The edge connectivity of the graph is 1
Single Difficult Graph  72. What is the number of edges present in a complete 1
choice algorithm, graph having n vertices?
 features A. (n*(n+1))/2
B. (n*(n-1))/2
C. n
D. Information given is insufficient
Single Difficult Graph 73. The given Graph is regular. 1
choice algorithm,
features

A. True
B. False
Single Difficult Graph 74. Consider the given graph. 1
choice algorithm,
Prims
algorithm

What is the weight of the minimum spanning tree

using the Prim’s algorithm, starting from vertex a?
A. 23
B. 28
C. 27
D. 11
Single Difficult Graph 75. Worst case is the worst case time complexity of 1
choice algorithm, Prim’s algorithm if adjacency matrix is used?
Prims A. O(log V)
algorithm B. O(V2)
C. O(E2)
D. O(V log E)
Single Difficult Graph  76. Which of the following is false about Prim’s 1
choice algorithm, algorithm?
Prims A. It is a greedy algorithm
algorithm B. It constructs MST by selecting edges in
increasing order of their weights
C. It never accepts cycles in the MST
D. It can be implemented using the Fibonacci heap
Single Medium Graph 77. What is the time complexity of Kruskal’s 1
choice algorithm, algorithm?
Kruskal A. O(log V)
algorithm B. O(E log V)
C. O(E2)
D. O(V log E)

Single Medium Graph 78. Which of the following is true? 1

choice algorithm, A. Prim’s algorithm can also be used for
Kruskal disconnected graphs
algorithm B. Kruskal’s algorithm can also run on the
disconnected graphs
C. Prim’s algorithm is simpler than Kruskal’s
algorithm
D. In Kruskal’s sort edges are added to MST in
decreasing order of their weights
Single Medium Graph 79. Consider the following statements. 1
choice algorithm, S1. Kruskal’s algorithm might produce a non-minimal
Kruskal spanning tree.
algorithm S2. Kruskal’s algorithm can efficiently implemented
using the disjoint-set data structure.
A. S1 is true but S2 is false
B. Both S1 and S2 are false
C. Both S1 and S2 are true
D.S2 is true but S1 is false
Single Medium Graph 80. Kruskal’s algorithm is best suited for the dense 1
choice algorithm, graphs than the prim’s algorithm.
Kruskal A. True
algorithm B. False

UNIT-V

Question Difficult Category, Questions Marks

Type y Subcategory
Single Easy Search 81. An algorithm which tries all the possibilities 1
choice algorithm, unless results are satisfactory is and generally is
DOP time consuming is:
A. Brute Force
B. Divide and Conquer
C. Dynamic programming algorithms
D. None of the mentioned

Single Medium Discrete 82. Optimization of algorithm means: 1

choice optimization A. making that algorithm fast by time and
problem, compact by space
introduction B. making that algorithm slow by time and large by
space
C. making that algorithm fast by time and large by
space
D. making that algorithm slow by time and compact
 by space

Single Medium Parallel 83. If W be serial work and WP be parallel work. 1

choice search , Search overhead factor s is defined as
Overhead A. W/Wp
B. Wp/W
C. p×(W/WP)
D. W*Wp

Single Easy Parallel DFS, 84. In Parallel Depth First search the Work is split 1
choice Work by
splitting A. Node Splitting
B. Stack Splitting
C. A & B both
D. None

Single Difficult Parallel Best 85. In Asynchronous Round Robin Scheme worst 1
choice first search, case is
splitting A. V(P)=O(P)
technique B. V(P)=O(P2)
C. V(P) is Unbounded
D. None of the above

Single Medium Parallel Best 86. Global round robin has poor performance 1
choice first search, because of
splitting A. large number of work requests.
technique B. contention at counter
C. a & b both
D. None

Single Medium Parallel Best 87. In Tree based Termination Detection Scheme 1
choice first search, Termination is signaled when the weight at
Termination processor P0 becomes
Detection A. 1
B. 0
C. –ve
D. None

Single Medium Parallel 88. In Parallel Formulation of Depth First Search 1

choice search, Branch and Bound If a processor's current best
Parallel DFS solution path is worse than the globally best
solution path, only the Correctness is effected not
its efficiency?
A. True
B. False

Single Medium Parallel DFS, 89. Drawback of common Cost Bound(Parallel 1

choice IDA* Formulation of IDA*) is
A. solution is not guaranteed to be optimal
B. there might not be enough
 concurrency
C. A & B both
D. Each processor works on a different
cost bound

Single Medium Parallel BFS, 90. In Parallel Best First Search the locking 1
choice operations operation is used to
A. extracts the best node
B. serialize queue access by various
processors
C. A & B both
D. None of the above

Single Medium Parallel BFS, 91. In Parallel Best First Search the run time will be 1
choice Execution at least
A. n(taccess+ texp)
B. (taccess+ texp)/taccess
C. ntaccess
D. None

Single Medium Parallel BFS. 92. Parallel Best first Search avoid contention by 1
choice Features A. balance the quality of nodes
B. balancing strategies
C. having multiple open lists.
D. A & B both

Single Medium Parallel BFS. 93. Parallel Best First Search balancing Strategies 1
choice Features are
A. Ring
B. Blackboard
C. Random Communication
D. All of above
Single Medium Parallel BFS. 94. In Parallel Best First Search Hashing can be 1
choice Features parallelized by
A. Two function
B. One function
C. Three function
D. None of the above

Single Medium Parallel BFS. 95. Executions yielding speedups greater than p by 1
choice Features using p processors are referred to as
A. deceleration anomalies
B. acceleration anomalies
C. Speedup anomalies
D. None of the above

Single Medium Parallel BFS. 96. If the heuristic function is good, the work done 1
choice Features in parallel best-first search is typically more than
 that in its serial counterpart.
A. True
B. False

Single Difficult Discrete 97. The objective of a DOP is to find a feasible 1

choice Optimization solution xopt, such that f(xopt) < f(x).
Problem, A. True
solution B. False
Single Medium Discrete 98. In DOP If the estimate is guaranteed to be an 1
choice Optimization underestimate, the heuristic is called
Problem, A. admissible heuristic
solution B. Optimal heuristic
C. A & B both
D. None of the above

Single Medium Discrete 99. Discrete Optimization Problems are generally 1

choice Optimization A. NP-Complete
Problem, B. NP-Hard
solution C. None of the above
D. Both A & B

Single Medium Sequential 100. DFBB does not explore paths that are 1
choice search, depth guaranteed to lead to solutions worse than current
first search best solution.
A. True
B. False

SHORT ANSWER TYPE QUESTIONS

UNIT-I

Question Difficult Category, Questions Marks

Type y Subcategory
Subjective Easy Implicit 1. Define Pipelining and Super Scalar 3
parallelism, execution with the help of an example.
Superscalar
execution
Subjective Medium Communication 2. What are the major differences between 3
model of message-passing and shared-address-
parallel space computers? Also outline the
platforms, advantages and disadvantages of the two.
shared address
space
Subjective Difficult Decomposition 3. A sample serial program to be parallelized. 3
technique, task 1. procedure FFT_like_pattern(A, n)
size 2. begin
3. m := log2 n;
 4. for j := 0 to m - 1 do
5. k := 2j;
6. for i := 0 to n - 1 do
7. A[i] := A[i] + A[i XOR 2j];
8. endfor
9. end FFT_like_pattern
In above algorithm, assume a decomposition such
that each execution of Line 7 is a task. Draw a
task-dependency graph and a task-interaction
graph.

Subjective Medium Dichotomy of 4. List three major problems requiring the use of 3
Parallel supercomputing in the following
Computing domains:
Platforms, 1. Structural Mechanics.
types 2. Computational Biology.
3. Commercial Applications.
Subjective Medium Routing 5. Why E-cube routing in a hypercube network is 3
mechanism, E- used? Explain with the help of an example.
cube routing in
hypercube
network
Subjective Easy Decomposition 6. Compare Recursive decomposition and Data 3
technique, decomposition techniques with a suitable
Recursive example.
decomposition
Subjective Easy Communication 7. Tasks may communicate with each other in 3
techniques, various ways. Define any two task
types communication technique with example.

Subjective Medium Parallel 8. Parallel computing has made a tremendous 3

programming impact on a variety of areas ranging from
platform, computational simulations for scientific and
features engineering applications to commercial
applications in data mining and transaction
processing. Present a small sample of the diverse
applications of parallel computing.
Subjective Easy Implicit 9. Differentiate following dependencies 3
parallelism a. True Dependency
,pipelining b. Resource dependency
superscalar c. Branch dependency
execution d. Procedural dependency
Subjective Difficult Communication 10. The time taken to communicate a message 3
Costs in between two nodes in a network is the sum of the
parallel time to prepare a message for transmission and
Machines, the time taken by the message to traverse the
Message network to its destination. Define the principal
Passing Costs parameters that determine the communication
in Parallel latency.
Computers
UNIT-II

Question Difficult Category, Questions Marks

Type y Subcategory
Subjectiv Difficult CUDA, 11. In CUDA a block can be split into parallel 3
e Blocks and threads. How to combine blocks and threads
threads together? Explain with the help of an example.
Subjectiv Easy CUDA 12. Explain CUDA Programming Model and 3
e architecture, Thread Batching.
CUDA
features
Subjectiv Medium CUDA, 13. When Threads are created? How kernel 3
e Kernel functions are called?
function
Subjectiv Medium CUDA, Block 14. Specify 1D grid and 2D block with the help of 3
e Grid an example.
dimensions
Subjectiv Medium CUDA 15. Demonstrate basic features of memory and 3
e parallel thread management in CUDA programs.
programming
model,
features
Subjectiv Medium CUDA, API 16. How to access Global variables via the Runtime 3
e functions API? Explain with a suitable example.
Subjectiv Medium CUDA, 17. How to share data between Threads? Explain 3
e Shared with suitable example.
memory

Subjectiv Easy CUDA 18. Explain Wrap based Thread scheduling? 3

e ,Thread
management
Subjectiv Difficult CUDA, 19. All threads in a grid execute the same kernel 3
e Thread function, they rely on unique coordinates to
Organization distinguish themselves from each other and to
identify the appropriate portion of the data to
process. How these threads are organized in Two
level hierarchy?
Subjectiv Easy CUDA, 20. Define the main functionality of bockId and 3
e Thread threadId.
Organization

UNIT-III

Question Difficulty Category, Questions Marks

Type Subcategory
Subjectiv Easy Analytical 21. Explain sources of overhead in Parallel 3
e modeling, programs.
Overheads
Subjectiv Medium Analytical 22. How to calculate the Total Overhead incurred 3
e modeling, by a Parallel program?
overhead
Subjectiv Difficult Analytical 23. Define superlinearity effects due to 3
e modeling, exploratory decomposition with an example.
Performance
metrics of
parallel
system
Subjectiv Difficult Analytical 24. Performance increased by increasing the 3
e modeling, granularity of computation in each processor. Prove
Granularity this statement with the help of an example.
and
performance
Subjectiv Medium Analytical 25. If a problem of size W has a serial component 3
e modeling, WS, prove that W/WS is an upper bound on its
Amdahl’s law speedup, no matter how many processing elements
are used.
Subjectiv Difficult Analytical 26. Can a parallel system keep efficiency by 3
e modeling, increasing the number of processors and the
scalability problem size simultaneously? Explain this
statement with example.
Subjectiv Easy Optimality, 27. Why isoefficiency function is required? Define 3
e isoefficiency isoefficiency Metric of scalability.
Subjectiv Medium Analytical 28. Find Minimum Cost‐Optimal Execution time 3
e modeling, for Adding n Numbers.
Cost optimal
Subjectiv Difficult Analytical 29. Why is performance extrapolation so difficult? 3
e modeling, Explain with the help of an example.
scalability of
parallel
system
Subjectiv Medium Analytical 30. Represent a cost-optimal way of computing the 3
e modeling, sum of 16 numbers using four processing elements.
Granularity

UNIT –IV

Question Difficult Category, Questions Marks

Type y Subcategory
Subjectiv Difficult Dense Matrix, Matrix 31. Write a serial algorithm for multiplying an 3
e vector multiplication n x n matrix A with an n x 1 vector x to yield
an n x 1 product vector y.
Subjectiv Difficult Dense Matrix, Matrix 32. Consider the two algorithms for all-to-all 3
e vector multiplication personalized communication. Which method
would you use on a 64-node parallel computer
with Q (p ) bisection width for transposing a
1024 x 1024 matrix with the 1-D partitioning
if ts = 100μ s and tw = 1μ s? Why?
Subjectiv Difficult Parallel computing 33. In the parallel formulations of bitonic sort, 3
e sorting, bitonic sort we assumed that we had n processes available
to sort n items. Show how the algorithm needs
to be modified when only n /2 processes are
available?
Subjectiv Medium Parallel computing  34. Difference between Comparison and Non- 3
e sorting, Sorting Comparison Sorting Algorithms?
Subjectiv Easy Minimum cost 35. Write down Prim's sequential minimum 3
e spanning tree, prims spanning tree algorithm.
Subjectiv Difficult Graph, BFS 36. Given a graph G = (V, E), the breadth-first 3
e ranking of vertices of G are the values
assigned to the vertices of V in a breadth-first
traversal of G from a node v. Show how the
breadth-first ranking of vertices of G can be
performed on a p-process mesh.
Subjectiv Difficult Connected 37. Compute the parallel run time, speedup, 3
e components, All pair and efficiency of Floyd's all-pairs shortest
shortest paths algorithm using 2-D block mapping on a
p-process mesh with store-and-forward routing
and a p-process hypercube and a p-process
mesh with cut-through routing.
Subjectiv Easy Sorting on parallel 38. Differentiate between internal and external 3
e computer, sorting algorithm?
introduction

Subjectiv Easy Sorting on parallel 39. How is a merge sort algorithm 3

e computers , merge implemented?
sort
Subjectiv Medium Graph algorithm , all 40. Which is the best parallel implementation 3
e pair shortest path to find single source shortest paths for graph
analytics?

UNIT V

Question Difficult Category, Questions Marks

Type y Subcategory
Subjectiv Easy Discrete 41. Define Discrete Optimization problem 3
e optimization, Basics with an example.
Subjectiv Easy Sequential search, 42. Define Depth first search by using IDA* 3
e IDA* technique.
Subjectiv Easy Sequential search, 43. Differentiate between BFS and DFS. 3
e BFS
Subjectiv Medium Parallel search , 44. Let W be the amount of work done by a 3
e Search Overhead single processor, and Wp be the total amount
 factor of work done by p processors. Then how to
define the search overhead factor?
Subjectiv Medium Parallel DFS, load 45. How to balance the load Dynamically in 3
e balancing Parallel Depth First search?
Subjectiv Medium Parallel DFS, Load 46. Differentiate between ARR,GRR schemes 3
e balancing with a suitable example.
Subjectiv Difficult Parallel DFS, A 47. Analyze the performance and scalability 3
e general framework of Parallel DFS algorithm for any Load
for analysis of balancing scheme with an example.
parallel DFS
Subjectiv Medium Searching, 48. Why Termination Detection is used? 3
e Termination Differentiate between Tree based termination
Detection detection and Dijkstra’s Termination
Detection techniques.
Subjectiv Easy Parallel BFS, 49. Why we use Blackboard Communication 3
e Communication strategy? Explain with the help of an
strategies example.
Subjectiv Medium Speedup anomalies, 50. Is it necessary to introduce Heuristic in 3
e Role of Heuristic Parallel Search? If yes, then explain the
reason.

LONG ANSWER TYPE QUESTION

UNIT-I

Question Difficult Category, Questions Marks

Type y Subcategory
Subjectiv Medium Memory 1. How to improve the effective memory 5
e performance, latency using cache? And explain the
caches impact of cache on memory system
performance.
Subjectiv Medium Memory 2. Consider a pipeline having 4 phases 5
e performance, cache with duration 60, 50, 90 and 80 ns.
Given latch delay is 10 ns. Calculate-
1. Pipeline cycle time
2. Non-pipeline execution time
3. Speed up ratio
4. Pipeline time for 1000 tasks
5. Sequential time for 1000 tasks
 6. Throughput

Subjectiv Difficult Improving 3. Consider a memory system with a level 5

e Effective Memory 1 cache of 32 KB and DRAM of 512
Latency Using MB with the processor operating at 1
Caches, Effective GHz. The latency to L1 cache is one
memory bandwidth cycle and the latency to DRAM is 100
cycles. In each memory cycle, the
processor fetches four words (cache line
size is four words). What is the peak
achievable performance of a dot product
of two vectors?
Subjectiv Difficult Physical 4. In how many subclasses PRAMs can be 5
e organization of subdivided? Of the four PRAM models,
parallel platform, which model is the most powerful?
architecture of Why?
Ideal parallel
computer
Subjectiv Difficult Routing in 5. Let A and B be two nodes in a d- 5
e interconnection, E- dimensional hypercube. Define H(A, B)
cube to be the Hamming distance between A
and B, and P(A, B) to be the number of
distinct paths connecting A and B. These
paths are called parallel paths and have
no common nodes other than A and B.
Prove the following:
1. The minimum distance in terms of
communication links between A and B is
given by H(A, B).
2. The total number of parallel paths
between any two nodes is P(A, B) = d .
3. The number of parallel paths between A
and B of length H(A, B) is
Plength=H(A,B)(A, B)= H(A, B).
4. The length of the remaining d - H(A, B)
parallel paths is H(A, B) + 2.

UNIT –II

Question Difficult Category, Questions Marks

Type y Subcategory
Subjectiv Medium CUDA, Thread 6. Define CUDA Thread Organization with 5
e Memory example. Also illustrates the organization of
Representation threads within a block.
Subjectiv Medium CUDA, 7. CUDA allows threads in the same block to 5
e Transparent coordinate their activities using a barrier
scalability synchronization function. Define this
synchronization function. How Transparent
 scalability in CUDA programs activated.
Subjectiv Easy CUDA, Device 8. Define CUDA API Functions for Device 5
e memories and data Global
transfer Memory Management and overview of CUDA
device Memory Model.
Subjectiv Difficult CUDA 9. How kernel function and Threading are 5
e programming related? Express this relation with the help of
model, Kernel Matrix Multiplication kernel Function in detail.
unction and
Threading
Subjectiv Medium CUDA 10. Explain Matrix multiplication (1D) with 5
e Programming, suitable example having 1 grid, 2 blocks, 2
Matrix threads.
multiplication

UNIT-III

Question Difficult Category, Questions Marks

Type y Subcategory
Subjectiv Difficult Analytical 11. Consider a parallel system containing p 5
e Modeling, processing elements solving a problem
Isoefficiency consisting of W units of work. Prove that if the
isoefficiency function of the system is worse
(greater) than Q(p), then the problem cannot be
solved cost-optimally with p = (W). Also prove
the converse that if the problem can be solved
cost-optimally only for p < Q(W), then the
isoefficiency function of the parallel system is
worse than linear.
Subjectiv Medium Analytical 12. What are the effects of concurrency on 5
e modeling, degree isoefficiency function? Explain detail with
of concurrency suitable example.
Subjectiv Difficult Analytical 13. What is Minimum Execution Time and 5
e modeling, Minimum Cost-Optimal Execution Time. How
minimum to find Minimum execution time for adding n
execution cost numbers?
Subjectiv Medium Analytical 14. Define all scaling characteristics of parallel 5
e modeling, scaling program with speedup and efficiency as fraction
characteristics of number of processing elements.
Subjectiv Easy Analytical 15.Explain all performance metrics for Parallel 5
e modeling, system.
performance
metrics
UNIT-IV

Question Difficult Category, Questions Marks

Type y Subcategory
Subjectiv Medium Matrix-matrix 16. Define a parallel algorithm for multiplying 5
e multiplication, NxN dense square matrices A and B to yield the
 square matrix product matrix C = A x B.
algorithm
Subjectiv Difficult Sorting, quick sort 17. Compare both sequential quick sort 5
e algorithm and parallelized quick sort algorithm
with an example.

Subjectiv Medium Parallel 18. Define shared address space parallel 5

e formulation, formulation and Message passing parallel
communication formulation.
Subjectiv Medium Other sorting 19. Write an algorithm for Enumeration sort on a 5
e algorithm, CRCW PRAM with additive-write conflict
Enumeration sort resolution.
Subjectiv Difficult Algorithms for 20. What are the sparse graphs? How to find a 5
e sparse graphs, Maximal Independent set? Explain with suitable
Maximal example.
independent set
UNIT-V

Question Difficult Category, Questions Marks

Type y Subcategory
Subjectiv Easy Sequential search, 21.Represent all algorithms that are based upon 5
e depth first search depth first search( Sequential Search) algorithm.

Subjectiv Difficult Analysis of parallel 22. Compute the value of V (p) for different 5
e DFS, Load load-balancing (ARR,GRR,RP) schemes.
balancing
Subjectiv Easy Termination 23. Define Tree based termination detection 5
e detection, tree technique.
based
Subjectiv Medium Parallel BFS, 24. Draw a general schematic diagram for 5
e representation parallel best-first search using a centralized
strategy.
Subjectiv Medium Parallel BFS, 25. Differentiate between ring communication 5
e communication and blackboard communication strategy.
strategy