Networking Reference
In-Depth Information
volume market, noting the original Cray-1 [ 54 ] in 1977 set a goal of delivering one system per
quarter. The proprietary signaling and wire protocols (packet formats, etc.) have made traditional
supercomputers incompatible with other vendors. However, the emergence of 40 gigabit and 100
gigabit Ethernet, coupled with a common processor-network interface like PCIe-Gen3
16 has
the potential to convolve datacenters and supercomputing into “super datacenters” — what Barroso
and Hölzle refer to as warehouse-scale computers [ 9 ].
Going forward, highly scalable machines capable of a exa flop computation will require low-diameter
scalable networks. Moreover, reliability and scalability are inseparable. System designers need to focus
on building power-efficient systems without sacrificing reliability in the hardware, while application
programmers will need to accept that servers are vulnerable to faults in components such as the
processors, memory, network, disks, etc., including system software; hardware can and does fail, and
programmers need to focus on fault-aware applications that can detect the loss of a component and
still function. Power-efficiency and reliability are the two largest impediments to continued scaling.
When programmers from both the supercomputing and Internet (datacenter) communities embrace
these concepts, we will benefit from greater interoperability and converged programming models
and best practices.
Search MirCeyron ::

Custom Search