Traditionally, every server node in a cluster must be equipped with enough extra hardware resources - i.e. memory, storage, co-processing, and/or other computing sub-systems - that stand ready to kick in when peak computing requirements demand them. Such required extra hardware resources is called "over-provisioning". Although over-provisioning resources remain unused and idle most of the time, they are always powered-on, thus they constantly consume energy power and add considerably to the operational costs of the cluster, 24/7 and on an on-going basis.
Ideally, the ability of any cluster node CPU to tap into idle hardware resources located inside other cluster nodes would releave data center owners of the considerable investment costs (CapEx) penalty and the on-going power usage costs (OpEx) of over-provisioning resources for each and every node in the cluster. Can such attractive goal be met? Not necessarily with current generation Ethernet- and InfiniBand-based cluster platforms, in which the typical, software-level Remote DMA (RDMA) method for node-to-node data exchange cannot even come close to providing the low latency access to remote node resources that hardware virtualization requires, as software is intrinsically latency-burdened. Unlike RDMA-based platforms and with its global shared memory architecture coupled with its intrinsic low latency capabilities, HyperShare empowers every node CPU in the cluster (regardless of cluster scale) with hardware-level, fast access to any near or remote cluster node resources - i.e. memory, storage, accceleration, etc. - as if they were its own local ones. Essentially, achieving a true hardware-level cluster resource vitualization that we call "Virtualization 2.0", which, in turn, enables a drastic reduction - ultimately the total elimination - of hardware resource over-provisioning for every node and across the entire cluster.
@112 Watts of Power Saved for Every 256 GB of Memory Not Required
@84 Watts of Power Saved for Every Terabyte of Storage Not Needed
@300 Watts of Power Saved for Every GPU Unit Not Required
Resulting in Unprecedented Reduction of Both Cluster-Wide Investment Costs (CapEx) and on-going Operational Costs (OpEx)
Virtualization 2.0 costs cutting opportunties do not stop there. In fact, certain costly and power-hungry hardware resources like memory, storage, acceleration, etc. can also be consolidated into function-dedicated cluster nodes that can serve part of or all the nodes across the cluster. In other words, HyperShare delivers true Distributed Resource Computing capabilities to data center owners , allowing them to build the cost and performance super-optimized data center infrastructures of the future today.
True Distributed Resource Computing
Resulting in even more drastic CaEx and OpEx cost cutting, unachievable with present days networking technology