Polygon examination

Содержание и основные понятия

Back to the table of contents
Alterna · Chapter 7 · page 7.4
Polygon examination
Check the knowledge on parallel computations and supercomputers.
Three stages of verification:
Quest
(XP warm-up)
control test
(10 random tasks from 25 bank) and
Open task
.
All tasks correspond to the EGE 2026 format, and task number 22 is parallel calculations.
EGE No. 22
01
Conformity map
Which topics of chapter 7 cover the task of the EGE 22
Subject
Page
What the EGE 22 checks
Amdal Law
7.1
Calculation of acceleration during dissipation
Flynn classification
7.1
Definition of architecture type (SISD/SIMD/MIMD)
Conwayer
7.1
Tactical calculation: k + (n-1)
Multiprecision
7.1
Deadlock, race condition, mutecs
FLOPS, units
7.2
Transfer of units of productivity
Cluster, architecture
7.2
Supercomputer components
GPU vs CPU
7.2
When GPU is Effective
OpenMP / MPI
7.2
Different approaches to parallelism
02
Quest: warm-up
8 scenarios up to 80 XP
♪ Quest: Parallel world
XP:
0
/ 80
03
Test
10 random tasks from the bank 25 · EGE format 22
For downloading, you need a local server (fetch does not work with file://).
Answered:
0
/
10
04
Open task
Case: Optimizing the parallel programme
- Task: "Accelerating the program"
The program processes an array of 1,000,000 elements.
70 per cent of time - treatment of each element (can be decoupled)
20% time - sorting of result (may be partial)
10% of time - file reading and initialization ( strict sequence)
Available: server with 8-nuclear processor and a video map with 4,096 cores.
Job:
Compute a theoretical acceleration on eight CPU kernels.
Evaluation criteria (max. 4 points):
Amdal's score (1 point):
p = 0.9 (70 per cent + 20 per cent), S = 1/(0.1 + 0.98) = 1/0.2125 E 4.7×. Limit: 1/0.1 = 10×.
Use of GPU (1 point):
The processing of elements (70%) is an ideal task for the GPU (SIMD, data parallism).
Synchronization problems (1 point):
Race reconciliation when recording results, need for a barrier after processing, transmission of data to the CPU\GPU (PCIe bottleneck).
Recommendations for action (1 point):
Minimize the sequence, use CUDA/OpenCL for GPU, profile to search for bottlenecks, take overhead into account for data transfer.
🚀
What's next?
You've studied parallel computations and supercomputers.
🔢
Duplicate coding
Class 10 · EGE 4
Fano's condition, decoding, uneven codes.
📐
Numeral systems
Class 10 EEG 14
Translations, arithmetic, a sign of division.
⊕
Boolean algebra
Class 10 EEG 2, 15
Five operations, laws, truth tables, diagrams.