void Balancer::balance(Cluster &cluster, Decomposition& decomp,
const int kConfig) {
const int kCPUAndGPU = 0;
const int kCPUOnly = 1;
const int kGPUOnly = 2;
int blocks_per_node = 0;
// num cpu nodes
unsigned int total_nodes = cluster.getNumNodes();
size_t num_blocks = decomp.getNumSubDomains();
// initialize block directory
cluster.setNumBlocks(num_blocks);
if (kConfig != kCPUOnly) {
unsigned int num_gpus = 0;
for (unsigned int node_index = 0;
node_index < cluster.getNumNodes();
++node_index) {
num_gpus += cluster.getNode(node_index).getNumChildren();
}
if (kConfig == kGPUOnly) // gpu only
total_nodes = num_gpus;
else if (kConfig == kCPUAndGPU) // cpu and gpu
total_nodes += num_gpus;
}
blocks_per_node = ceil(decomp.getNumSubDomains() / (float) total_nodes);
for (unsigned int node_index = 0;
node_index < cluster.getNumNodes();
++node_index) {
Node& node = cluster.getNode(node_index);
if (kConfig == kCPUOnly || kConfig == kCPUAndGPU) {
for (int subd = 0;
subd < blocks_per_node && 0 < decomp.getNumSubDomains();
++subd) {
SubDomain *block = decomp.popSubDomain();
node.addSubDomain(block);
}
}
if (kConfig == kGPUOnly || kConfig == kCPUAndGPU) {
for (unsigned int gpu_index = 0;
gpu_index < node.getNumChildren();
++gpu_index) {
Node& gpu = node.getChild(gpu_index);
for (int subd = 0;
subd < blocks_per_node && 0 < decomp.getNumSubDomains();
++subd) {
SubDomain *block = decomp.popSubDomain();
gpu.addSubDomain(block);
}
}
}
}
/* the work is balanced, so we can fill the block directory */
cluster.storeBlockLocs();
}
