Load-balancing problem

Lindh, R., J. W. Krogh, M. Schutz, and K. Hirao. Semidirect parallel self-consistent field The load balancing problem in the input/output intensive self-consistent field iterations. Theor. Chem. Acc. 110 156-164,2003. 

Load-Balancing Problem A finite-volume representation of a partial differential equation can be solved most efficiently by a single-instruction multiple-data (SIMD) computer system if the grid is partioned approximately proportionally for each... 

This optimization problem for grid distribution is known as the load-balancing problem. Its solution requires the minimization of the cost function H = Tcomp +... 

As an illustration of the load balancing achieved in a typical 32 processor calculation, we present the CPU usage for each of the 31 slave processes in Fig. 4. The difference between maximum and minimum CPU times is here 0.41 sec., and the overall root mean square deviation is 0.45%. This is very satisfactory, and it would seem unlikely that load balancing will present serious problems in future applications unless a significantly larger number of processors is employed. 

Our previous conclusions will all have some dependence on the particular type of application considered, an aspect which may influence the total CPU work load, the amount of data transfer required, and the possibility of obtaining even load balancing. Our findings in this work would suggest that load balancing does not constitute a problem for Cl applications of reasonable size, and the success or failure of the parallelization will therefore be determined... 

Off-center errors are problems specifically associated with top-loading balances. Placing a balance pan above the fulcrum places different torques and friction on balance pieces that do not exist when a balance pan hangs. The problem is exhibited if an object has different weight readings when moved to various locations across the surface of a top-loading balance pan. 

Sato, Tanaka, and Yao also reported a parallel implementation of the AMBER MD module.The target machine, the APIOOO distributed-memory parallel computer developed at Fujitsu, consisted of up to 1024 processor elements connected with three different networks. To obtain a higher degree of parallelism and better load balance between processors, a particle division method was developed to randomly allocate particles to processors. Experiments showed that a problem with 41,095 atoms can be processed 226 times faster with a 512-processor APIOOO than by a single processor. 

World-wide there were developed and implemented eontrol structures that solve particular problems of load balancing control [4, 5,6]. 

Parallel computers can be divided into two classes, based on whether the processors in the system have their own private memory or share a common memory. In a distributed memory system, the processors communicate with each other by sending and receiving messages through a communication network connecting all the processors. The problem to be solved must be explicitly partitioned by the programmer onto the various processors in such a way that load balancing is maintained and communication between processors is minimized and well ordered. For some problems it may not be easy or even... 

The architecture, depicted in Fig. 4.12, reflects the tasks that are performed by the management platform. Task distribution, performed primarily by the service trader and load balancer (see Subsect. 4.2.5), plays a central role here. Configuration, performance, fault, and security management are separate entities and are described in Subsect. 4.2.6. The problem of the integration of legacy tools is described in Subsect. 4.2.4. 

The resulting winner determination problem was a set-partitioning problem. Cost was not the only criterion. Carrier-reliability, load balancing and incumbency concerns mattered to THD. These nonmonetary preferences were not revealed to the bidders. 

Even if the algorithm can guarantee the correct result, load imbalance may remain a problem. This is particularly tme when static load balancing has been employed, since this technique typically assumes that all processors will do a particular job in the same amount of time. Static load balancing can still be used, but to obtain satisfactory results, accurate information about the relative speeds of the processors for doing the task at hand will be needed. 

A fundamental issue in the design of parallel algorithms is how to divide the computational problem into a set of tasks to be distributed among processors. Not only should the distribution of tasks provide load balance for a given number of processors (i.e., ensure that all processors are fully occupied all the time), but also the definition of tasks should be sufficiently flexible to provide a load-balanced execution over a widely varying number of processors. 

---