As manycore system accommodates multiple guest OSes and tasks inside them, resource
management become a key issue. More concurrent tasks increases needs for
larger and faster memory hierarchy, memory wall decreases benefit from the concurrent
executions of tasks, especially in the virtualized environment. To provide enough
computing resources for consolidated VMs in a energy-efficient manner, we targets
a manycore system with a memory hierarchy of on-chip DRAM, off-chip DRAM and
off-chip NVRAM. In this article, we provided the MN-MATE, an energy-efficient, fullfledged
resource management system for consolidated VMs in a single hardware node.
We proposed a runtime placement policy of consolidated VM and a memory balancing
scheme among them in the hypervisor. On top of this, We build two guest OS kernels,
Linux and our own lightweight kernel named CORE kernel. With the balanced CPU
and memory resources, we proposed a resource-aware task scheduling algorithm, polymorphic
management scheme of on-chip/off-chip DRAMs, and data placements and
management policies for hybrid main memory of off-chip M2 DRAM and off-chip M2
NVRAM for the Linux guest OS. We also build a simplified kernel to accelerate execution
of scientific applications through NUMA-aware buffer selectionmechanism. Consequently,
MN-MATE significantly enhanced system performance with the new hardware
system with an improved energy efficiency.We can see performance improvement with
our applications, core federation and QR factorization. The experiment results show
that MN-MATE outperforms other competing resource management schemes in terms
of application execution time and energy consumption.