Page 125 - Computer_Science_F5

Computer Science Data Level Parallelism (d) DLP can contribute to reduced power
consumption by potentially enabling
While TLP focuses on dividing the task
faster execution times. By completing
itself, DLP focuses on dividing the data.
tasks quicker, processors can spend
For instance, if the task is sorting candy,
less time in active states and more
DLP would focus on dividing the candy.
FOR ONLINE READING ONLY
Each friend could grab a handful of candy
finding the right balance between
and sort them by color independently. In time in low-power modes. However,
the context of the pizza party, DLP might parallelism and energy efficiency
involve dividing the task of decorating requires careful optimisation.
the room further. One friend could hang
streamers concurrently while another Challenges of DLP
puts up balloons, effectively utilising (a) Data Dependencies: DLP may not
multiple “cores” (people) to complete be suitable for tasks with high data
the task faster. Data-Level Parallelism dependencies, where processing one
(DLP) is a technique for improving data element relies on the results of
program performance by distributing another.
operations on large datasets across (b) Granularity: Finding the optimal data
multiple processing units. Instead of chunk size for efficient processing can
focusing on dividing the task itself (like be crucial. Too small chunks can lead
in TLP), DLP emphasises splitting the to increased overhead, while overly
data into smaller chunks and processing large chunks might not utilise all
them simultaneously on different cores processing units effectively.
or processors. (c) Algorithm design: Implementing

data-level parallelism requires careful
Advantages offered by DLP algorithm design to ensure proper
(a) Improved performance: Dividing data division and synchronization
the workload amongst multiple between processing units.
processing units can significantly
reduce execution time for data-
intensive tasks. Applications of DLP

(b) Scalability: DLP applications (a) Image processing: Consider an
can potentially scale well with application that needs to apply a filter
increasing hardware resources (cores, (Such as, blur and sharpen) to a high-
processors) by further dividing the resolution image. DLP allows us to
data for parallel processing. divide the image into smaller tiles.
Each core can then independently
(c) Reduced overhead: Compared to TLP, apply the filter to its assigned
DLP often incurs less overhead for tile, significantly accelerating the
context switching and maintaining processing time compared to a single-
separate thread states. core approach.

116
for Advanced Secondary Schools

Computer Science Form 5.indd 116 23/07/2024 12:33