2023 10th International Conference on Advanced Materials and Computer Science (ICAMCS 2023)
Research on the Application of Artificial Intelligence Algorithms in GPU
Resource Optimization and Scheduling Control
Zhen Li
Information Engineering Department, Jiangsu Maritime Institute, Nanjing, 211100, China
Keywords: Artificial intelligence algorithms; GPU; Resource optimization and scheduling; Control;
Application
Abstract: GPU is a new commercial computing model and an extension and expansion of
technologies such as distributed parallel processing and grid computing. It represents a new stage in
the development of parallel computing technology and belongs to a brand new product. With the
maturity of this field and the continuous growth of the internet and global enterprises, it is expected
that more cloud service providers will emerge. This article focuses on the research and development
scenarios of AI(Artificial Intelligence) algorithms and conducts a series of scheduling strategy
studies based on GPU to improve the task execution efficiency and cluster resource utilization of
the AI algorithm development platform. Finally, the research results indicate that when the
historical time slot changes from 50 to 590, the accuracy of prediction becomes more and more
accurate as the size of the time slot increases, and the prediction curve gradually becomes flat at 410
to 460. It can be inferred that there may be an extreme value around the time slot size of 460, and
exceeding this extreme value has little impact on predicting low utilization time results. During the
process of increasing average execution power, resource correlation algorithms are used to schedule
tasks, reducing the overhead of the cloud system data center network and improving the system's
resource utilization.
1. Introduction
In recent years, GPU has made breakthrough achievements in the fields of floating-point
computing and parallel computing, and has great potential for improvement. GPU is a brand-new
commercial computing mode, and it is also the extension and expansion of distributed parallel
processing and grid computing. It represents a brand-new stage of the development of parallel
computing technology and belongs to a brand-new product. There are many problems in GPU that
need to be solved effectively, such as resource scheduling. The heterogeneous computing system
based on GPU can provide computing power to meet the needs of industrial manufacturing, AI,
medical care, aerospace and other fields under reasonable algorithm scheduling, and can provide
core technologies for China's economic development and provide strong support for the strategy of
"strengthening the country through science and technology" [1]. Resource scheduling is the main
component of GPU, and its efficiency has a direct impact on the performance of GPU environment.
Heuristic intelligent algorithm is the main research direction in this field. With the maturity of this
field and the continuous growth of Internet and enterprises in the world, it is expected that more
cloud service providers will emerge, provide a more diversified choice, more resources and services,
and increase energy consumption and negative pressure on the environment [2-3]. However, the
increased network circulation load and energy consumption brought by the network are gradually
increasing, and the energy consumption of GPU is rarely taken into account in the existing resource
scheduling algorithms [4]. In resource scheduling control, the energy consumption of GPU is still
rarely involved. Although GPU can provide computing and storage services to save energy by using
large shared servers and storage units, this effect will be especially obvious when end users tend to
use a computer or terminal with low capacity and reduce energy consumption migration [5]. How to
improve the utilization of GPU resources, and then improve the task throughput of AI algorithms,
has gradually attracted more and more attention. In this paper, a series of scheduling strategies are
studied based on GPU to improve the task execution efficiency and cluster resource utilization of
Published by CSP © 2023 the Authors 1
the AI algorithm research and development platform.
2. Research on GPU Resource Optimization and Scheduling Control Based on AI Algorithms
2.1. System control algorithm
The productization research and development of AI from academic research to production
applications usually involves various steps and tools, especially in environmental deployment,
testing, tuning, and other work, which often consumes a lot of time and expenses. Usually, when
optimizing resource scheduling, each task is matched with the best computing unit, and in task
replication based scheduling algorithms, the same computing task can be assigned to multiple
computing units. The algorithm defaults to no communication overhead between tasks located in
the same computing unit, By copying the predecessor task of a task to the idle time slice of its
computing unit, the interaction cost between tasks and the idle time of the computing unit are
reduced [6-7]. This algorithm uses the computational cost of computing units to exchange for
communication costs between tasks, and is suitable for computing tasks with frequent
communication between nodes. In order to make the controller adaptive and achieve automatic
adjustment of controller parameters, AI algorithms can be used. Due to the parallel mechanism,
learning ability, and memory function of AI algorithms, utilizing the self-learning characteristics of
AI algorithms and combining traditional ideas, a GPU resource optimization scheduling controller
is constructed, achieving automatic adjustment of conventional controller parameters [8]. The
principle of the control system is shown in Figure 1.
Figure 1 Control System Schematic Diagram
The core concept is to construct a critical path task sequence based on b-level and t-level values,
and then copy the GPU and its direct predecessor tasks to the same computing unit, advancing the
start time of the task. The calculation formula is as follows:
t1(ti ) = max (ci , j + w(t j ) + tl (ti ))
(1)
b1(ti ) = w(ti ) + max (ci , j + bl (t j ))
(2)
The core concept of the algorithm is to execute Join tasks as early as possible. The Fork Join
graph dedicated algorithm addresses the issue of interactive channel conflicts in the actual GPU
resource optimization scheduling process. The controller provides a flexible framework for the
construction of control strategies. By adjusting parameters, conflicts between various performance
indicators in the control system can be effectively reduced [9-10].
2.2. Resource optimization scheduling control strategy
Our GPU resource optimization and scheduling strategy targets scenarios where data center tasks
arrive dynamically, with the goal of maximizing system throughput as much as possible. System
throughput can be measured by the number of instructions per cycle (IPC) executed in the kernel.
The expression for IPC is as follows:
2
Inst
IPC =
Cycles (3)
Among them, Inst represents the total number of running instructions for a kernel, Cycles
represents the number of clock cycles required for the kernel to run alone, and the IPC of the kernel
can be expressed as the ratio of the total number of running instructions for the kernel to the number
of clock cycles required for the kernel to run. We use normalized IPC to measure system throughput.
The normalized IPC of a kernel refers to the IPC that is normalized to the kernel when running
separately. The specific expression for normalized IPC is as follows:
IPC MK
norm _ IPC =
IPC SK (4)
MK
Among them, IPC represents the IPC exhibited by the kernel when running together with
SK
other kernels, and IPC represents the IPC exhibited by the kernel when running alone.
The ultimate goal of scheduling strategies is to maximize the overall system throughput in
typical application scenarios where data center tasks arrive dynamically. The tasks submitted by the
user will be imported into the task waiting queue, and during the scheduling process at each
scheduling point, the selection of placement strategies and the calculation of waiting margin will be
completed based on performance prediction models. At the same time, the tasks in the task queue
will be reordered based on the waiting margin. The implementation of waiting queues is based on a
heap structure, which can achieve automatic sorting of tasks based on event triggering mechanisms.
When the number of task nodes in a parallel task graph far exceeds the number of computing units
in a heterogeneous system, the communication overhead between subtasks during the execution
process is extremely high, which increases the computational cost. Aggregating small-scale tasks
together, reducing the number of tasks to be scheduled, and increasing the task size in exchange for
a reduction in communication overhead and scheduling length. Our scheduling strategy mainly
consists of two performance analysis modules: kernel collaborative operation performance analysis
module and kernel analysis module, as well as research on AI algorithms based on the above two
performance analysis modules in GPU resource optimization scheduling control.
3. Experimental results and analysis
3.1. Experimental setup
Due to the high utilization rate of GPU resource optimization, the task may fail due to too long
waiting time, or because of too long execution time, although all submitted tasks have been
completed, it wastes too much energy consumption of the cloud system. For such a large number of
tasks coming in a certain period of time, this chapter puts forward the energy consumption
optimization research of GPU resource scheduling in AI system based on task tolerance value. In
this chapter, in order to test the performance and efficiency of the energy consumption optimization
algorithm of cloud computing resource scheduling designed by time prediction in different
environments, a cloud computing test platform with 130 nodes is built by using Cloud Sim
simulation under the laboratory platform. In the experiment, in order to make the algorithm get
better performance and not lose too much overhead of the system, the following settings and
functions are verified. Firstly, the system resource utilization rate of the algorithm and the setting of
historical time slots are verified. Secondly, the experimental scale of the proposed algorithm and the
algorithm to be compared is set. When the cloud node is 130, the experiment of system utilization
threshold y and total energy consumption of the system is carried out, and the average value of 95
results is selected as the experimental result. The analysis is shown in Figure 2.
3
Figure 2 Schematic diagram of the relationship between system utilization threshold and energy
consumption
As shown in Figure 2, it can be seen that when the system utilization threshold y is between 0-
25%, the total energy consumption of the system does not change much, and frequent transitions to
the system's state consume too much energy, which is offset by the energy saved by sleep. When
the system utilization rate is between 35% -70%, the energy consumption of the system gradually
increases. After multiple experiments and repeated testing, it has been found that when the system
utilization threshold is about 75%, the system can achieve the best results through optimization
algorithms.
130 nodes were selected for an experiment on system prediction accuracy with a random task of
80 and a cloud system scale. The experimental results were taken as the average energy
consumption of 95 tests. The experimental analysis of historical time slots and accurate prediction
is shown in Figure 3.
Figure 3 Schematic diagram of time slot and accuracy
From Figure 3, it can be seen that when the historical time slot changes from 50 to 590, the
accuracy of the prediction becomes more and more accurate as the size of the time slot increases,
and the prediction curve gradually becomes flat at 410 to 460. It can be inferred that there may be
an extreme value around the time slot size of 460, and exceeding this extreme value has little impact
on predicting low utilization time results.
3.2. Experimental Summary
When the node size of the cloud system is 0-128 nodes, the system utilization threshold y is 72%,
4
and the historical time slot size is 570, the average energy consumption of the algorithm based on
time prediction is about 7.62% lower than that of the average prediction algorithm under different
task sizes, while it is 20.52% lower than that of the management optimization strategy. With the
gradual increase of tasks, the algorithm based on time prediction gradually approaches and equals to
the idle energy consumption of management optimization algorithm. Achieve higher service
availability through network redundancy. Although only one data center is designed, cloud service
providers usually maintain redundant centers for efficient data transportation and load balancing
among these data centers. The goal of virtualization is to keep tasks parallel, improve flexibility,
scalability and utilization of hardware resources, and then reduce the hardware and operating costs
of data centers. Random task sets are scheduled to cloud service resource nodes for execution
through task waiting queues. However, the arrival time of random tasks is uncertain and the
resource requirements are unequal. The research assumes that the task waiting queue is long enough,
and the cloud resource service center is responsible for reasonably scheduling task execution. The
biggest advantage of adopting AI technology is that it can flexibly map the performance strength of
physical resources to virtual servers. By appropriately reducing the resources allocated to tasks, the
system load rate is increased to optimize the resource utilization rate of cloud computing nodes. By
increasing the parallelism of task execution, the task waiting time is reduced, and then the waiting
energy consumption is optimized. Historical time slot algorithm can reduce the error rate of
prediction and adjust it quickly when there is an error; The proposed algorithm saves more energy
than the average energy consumption prediction algorithm and energy consumption management
optimization algorithm by setting the system utilization threshold and selecting the time slot size,
and also shows some advantages in some system performance.
4. Conclusions
In order to reduce the transmission energy consumption during the execution of cloud computing
systems when the random task volume is normal and moderate, this article designs an optimized
scheduling algorithm based on cloud computing resource correlation energy consumption from the
perspective of resource correlation. This article conducts research on the application of AI
algorithms in GPU resource optimization scheduling control. This article also proposes a resource
optimization scheduling control strategy. When the number of task nodes in the parallel task graph
far exceeds the number of computing units in heterogeneous systems, the communication overhead
between subtasks during the calculation execution process is extremely high, which increases the
computational cost. Aggregating small-scale tasks together, reducing the number of tasks to be
scheduled, and increasing the task size in exchange for a reduction in communication overhead and
scheduling length. The optimization of transmission energy consumption was achieved through task
scheduling and allocation processes in AI algorithms, as well as network load levels. The research
results indicate that when the historical time slot changes from 50 to 590, the accuracy of prediction
becomes more and more accurate as the size of the time slot increases, and the prediction curve
gradually becomes flat at 410 to 460. It can be inferred that there may be an extreme value around
the time slot size of 460, and exceeding this extreme value has little impact on predicting low
utilization time results. During the process of increasing average execution power, resource
correlation algorithms are used to schedule tasks, reducing the overhead of the cloud system data
center network and improving the system's resource utilization. This article also designs a resource
optimization scheduling control strategy to achieve a balance between the allocation of memory
resources and task execution performance.
References
[1] Nguyen T T, Vo D N.The application of an effective cuckoo search algorithm for optimal
scheduling of hydrothermal system considering transmission constraints[J].Neural Computing &
Applications, 2022, 18(9):1-22.
5
[2] Nguyen T T, Vo D N.The application of an effective cuckoo search algorithm for optimal
scheduling of hydrothermal system considering transmission constraints[J].Springer London, 2019,
30(8):10-18.
[3] Kang L, Liu D, Wu Y,et al.An Improved DE Algorithm for Solving Multi-Furnace Optimal
Scheduling of Single Crystal Silicon Production[J].International Journal of Pattern Recognition and
AI, 2023, 37(02):10-22.
[4] Shi H, Cao G, Ma G,et al.Intelligent Service Scheduling Decision Technology under AI
Technology[J].Journal of Physics: Conference Series, 2021, 1986(1):012078-012087.
[5] Patronas G, Vlassopoulos N, Bellos P,et al.Accelerating the scheduling of the network resources
of the next-generation optical data centers[J].Parallel computing, 2023, 19(20):36-44.
[6] Meng R, Rao Y, Zheng Y,et al.Modelling and solving algorithm for two-stage scheduling of
construction component manufacturing with machining and welding process[J].International
Journal of Production Research, 2018, 56(19):6378-6390.
[7] Liu D, Yao Z, Chen L.Emergency Scheduling Optimization Simulation of Cloud Computing
Platform Network Public Resources[J].Complexity, 2021, 2021, 20(11):10-18.
[8] Jordehi A R.Enhanced leader particle swarm optimisation (ELPSO): a new algorithm for
optimal scheduling of home appliances in demand response programs[J].AI Review, 2020,
53(3):2043-2073.
[9] Wang Y, Qiao J, Lin S,et al.An Approximate Optimal Solution to GPU Workload
Scheduling[J].Computing in Science & Engineering, 2018, PP(5):21-30.
[10] Liu S, Wang N.Collaborative Optimization Scheduling of Cloud Service Resources Based on
Improved Genetic Algorithm[J].IEEE Access, 2020, 86(99):11-19.
