In computing, load balancing is the process of distributing a set of tasks over a set of resources (computing units) with the aim of making their overall processing more efficient. Load balancing can optimize response time and avoid unevenly overloading some compute nodes while other compute nodes are left idle.
Load balancing is the subject of research in the field of parallel computers. Two main approaches exist: static algorithms, which do not take into account the state of the different machines, and dynamic algorithms, which are usually more general and more efficient but require exchanges of information between the different computing units, at the risk of a loss of efficiency.
Problem overview A load-balancing algorithm always tries to answer a specific problem. Among other things, the nature of the tasks, the algorithmic complexity, the hardware architecture on which the algorithms will run as well as required error tolerance, must be taken into account. Therefore compromise must be found to best meet application-specific requirements.
Nature of tasks The efficiency of load balancing algorithms critically depends on the nature of the tasks. Therefore, the more information about the tasks is available at the time of decision making, the greater the potential for optimization.
Size of tasks Perfect knowledge of the execution time of each of the tasks allows to reach an optimal load distribution (see algorithm of prefix sum). Unfortunately, this is in fact an idealized case. Knowing the exact execution time of each task is an extremely rare situation.
For this reason, there are several techniques to get an idea of the different execution times. First of all, when tasks are relatively homogeneous size, it is possible to consider that each of them will require approximately the average execution time. If, on the other hand, the execution time is considerably irregular, other techniques may be used. One technique is to add some metadata to each task. Depending on the previous execution time for similar metadata, it is possible to make inferences for a future task based on statistics.
Dependencies In some cases, tasks depend on each other. These interdependencies can be illustrated by a directed acyclic graph. Intuitively, some tasks cannot begin until others are completed.
Assuming that the required time for each of the tasks is known in advance, an optimal execution order must lead to the minimization of the total execution time. However, this is an NP-hard problem and therefore can be difficult to solve exactly. There are algorithms, like job scheduler, that calculate optimal task distributions using metaheuristic methods.
Segregation of tasks Another feature of the tasks critical for the design of a load balancing algorithm is their ability to be broken down into subtasks during execution. The tree-shaped computation algorithm presented later takes great advantage of this specificity.
Static and dynamic algorithms
Static A load balancing algorithm is "static" when it does not take into account the state of the system for the distribution of tasks. Thereby, the system state includes measures such as the load level (and sometimes even overload) of certain processors. Instead, assumptions about the overall system are made beforehand, such as the arrival times and resource requirements of incoming tasks. In addition, the number of processors, their respective power and communication speeds are known. Therefore, static load balancing aims to associate a known set of tasks with the available processors in order to minimize a certain performance function. The key design choice is the performance function.
Static load balancing techniques are commonly centralized around a router, or master, which distributes the loads and optimizes the performance function. This minimization can take into account information related to the tasks to be distributed, and derive an expected execution time.
The advantage of static algorithms is that they are easy to set up and can be efficient in the case of fairly regular tasks (such as processing HTTP requests from a website). However, there is still some statistical variance in the assignment of tasks which can lead to the overloading of some computing units.
Dynamic Unlike static load distribution algorithms, dynamic algorithms take into account the current load of each of the computing units (also called nodes) in the system. In this approach, tasks can be moved dynamically from an overloaded node to an underloaded node in order to receive faster processing. While these algorithms are much more complicated to design, they can improve performance, in particular, when the execution time varies greatly from one task to another.
Dynamic load balancing architecture can be more modular since it is not mandatory to have a specific node dedicated to the distribution of work. When tasks are uniquely assigned to a processor according to their state at a given moment, it is a unique assignment. If, on the other hand, the tasks can be permanently redistributed according to the state of the system and its evolution, this is called dynamic assignment. A load balancing algorithm that requires too much communication in order to reach its decisions runs the risk of slowing down the resolution of the overall problem.
Hardware architecture Heterogeneous machines Parallel computing infrastructures are often composed of units of different computing power, which should be taken into account for the load distribution.
For example, lower-powered units may receive requests that require a smaller amount of computation, or, in the case of homogeneous or unknown request sizes, receive fewer requests than larger units.
Shared and distributed memory Parallel computers are often divided into two broad categories: those where all processors share a single common memory on which they read and write in parallel (PRAM model), and those where each computing unit has its own memory (distributed memory model), and where information is exchanged by messages.
For shared-memory computers, managing write conflicts greatly slows down the speed of individual execution of each computing unit. However, they can work perfectly well in parallel. Conversely, in the case of message exchange, each of the processors can work at full speed. On the other hand, when it comes to collective message exchange, all processors are forced to wait for the slowest processors to start the communication phase.
In reality, few systems fall into exactly one of the categories. In general, the processors each have an internal memory to store the data needed for the next calculations and are organized in successive clusters. Often, these processing elements are then coordinated through distributed memory and message passing. Therefore, the load balancing algorithm should be uniquely adapted to a parallel architecture. Otherwise, there is a risk that the efficiency of parallel problem solving will be greatly reduced.
Hierarchy Adapting to the hardware structures seen above, there are two main categories of load balancing algorithms. On the one hand, the one where tasks are assigned by “master” and executed by “workers” who keep the master informed of the progress of their work, and the master can then take charge of assigning or reassigning the workload in case of the dynamic algorithm. The literature refers to this as master-worker architecture. On the other hand, the control can be distributed between the different nodes. The load balancing algorithm is then executed on each of them and the responsibility for assigning tasks (as well as re-assigning and splitting as appropriate) is shared. The last category assumes a dynamic load balancing algorithm.
Since the design of each load balancing algorithm is unique, the previous distinction must be qualified. Thus, it is also possible to have an intermediate strategy, with, for example, "master" nodes for each sub-cluster, which are themselves subject to a global "master". There are also multi-level organizations, with an alternation between master-slave and distributed control strategies. The latter strategies quickly become complex and are rarely encountered. Designers prefer algorithms that are easier to control.
Adaptation to larger architectures (scalability) In the context of algorithms that run over the very long term (servers, cloud...), the computer architecture evolves over time. However, it is preferable not to have to design a new algorithm each time.
One parameter of a load balancing algorithm is therefore its ability to adapt to scalable hardware architecture. This is called the scalability of the algorithm. An algorithm is called scalable for an input parameter when its performance remains relatively independent of the size of that parameter.
When the algorithm is capable of adapting to a varying number of computing units, but the number of computing units must be fixed before execution, it is called moldable. If, on the other hand, the algorithm is capable of dealing with a fluctuating amount of processors during its execution, the algorithm is said to be malleable. Most load balancing algorithms are at least moldable.
Fault tolerance Especially in large-scale computing clusters, it is not tolerable to execute a parallel algorithm that cannot withstand the failure of one single component. Therefore, fault tolerant algorithms are being developed which can detect outages of processors and recover the computation.
Approaches
Static distribution with full knowledge of the tasks: prefix sum If the tasks are independent of each other, and if their respective execution time and the tasks can be subdivided, an optimal optimal algorithm is available under these assumptions.
By dividing the tasks in such a way as to give the same amount of computation to each processor, all that remains to be done is to group the results together. Using a prefix sum algorithm, this division can be calculated in logarithmic time with respect to the number of processors.
If, however, the tasks cannot be subdivided (i.e., they are atomic), although optimizing task assignment is a difficult problem, it is still possible to approximate a relatively fair distribution of tasks, provided that the size of each of them is much smaller than the total computation performed by each of the nodes.
Most of the time, the execution time of a task is unknown and only rough approximations are available. This algorithm may be efficient under these assumptions, but is less applicable when task durations are unknown.
Static load distribution without prior knowledge Even if the execution time is not known in advance at all, static load distribution is always possible.
Round-robin scheduling In a round-robin algorithm, the first request is sent to the first server, then the next to the second, and so on down to the last. Then it is started again, assigning the next request to the first server, and so on.
This algorithm can be weighted such that the most powerful units receive the largest number of requests and receive them first.
Randomized static Randomized static load balancing is simply a matter of randomly assigning tasks to the different servers. This method works quite well. If, on the other hand, the number of tasks is known in advance, it is even more efficient to calculate a random permutation in advance. This avoids communication costs for each assignment. There is no longer a need for a distribution master because every processor knows what task is assigned to it. Even if the number of tasks is unknown, it is still possible to avoid communication with a pseudo-random assignment generation known to all processors.
The performance of this strategy (measured in total execution time for a given fixed set of tasks) decreases with the maximum size of the tasks.
Others There are other methods of assignment as well: