Future Big Systems

 August 24, 2011, Hot Interconnects 19, Santa Clara, CA—Greg Papadopoulos, venture partner at New Enterprise Associates and former CTO of Sun Microsystems talked about “The Computer That is The Network: The Future History of Big Systems”.

Hot interconnect 19 2011

We are in the midst of a revolution where the cost to connect drops by an order of magnitude over time. At the same time, there is an increasing number of things to connect. In the 90s, it cost about $100 to connect a computer to another one. In 2000, it cost about a dollar to connect people electronically, and in this decade, things will talk to each other for one cent.

Wide-area network bandwidth and last mile technologies are improving communications speeds, while wireless is also seen dramatic improvements. Now, many people can get from 10 to 100 M bits per second bandwidth for everyday use. Some of the surprises of his well-connected environment are the expectations from users are now lower than the reality. The advent of cloud computing and storage has created a cloud flair, and even include cloud-based firewalls.

One result of increasing bandwidth and number of devices is a transition to bigger computers with large compute centers going from 100,000 to over 1 million nodes. These transitions are also driving consolidation towards economies of scale. But the compute centers also need to be attached to a distribution network to be viable. Now, start-up companies delivering large data volumes can contract for cloud services and, if successful, later build their own networks.

We have seen very many cycles of computing starting from mainframes, to mini computers, PCs and now to massively parallel supercomputers. All of these systems are based on old components. Technology changes enable the next generation, and we will see SOC and multi-core designs integrate monolithic switches next. The cloud will move to exa-scale systems, and the creation of new systems will be left to the user. The reason systems design becomes the user’s responsibility is that some components are good but there are no integrated systems.

Even the operating systems are questionable. This results from our views of the systems. The classical view emphasized processing and resulted in clusters of small scale computers. Newer systems are using virtualization and management to improve operational efficiencies. Next, we will see network virtualization management with emphasis on the network. So it will move from a processing focus to a big data and storage operation with virtualization and management layers. We should see more PaaS as people develop new linguistics models

On top of the different views, there are also different operating principles. Older style systems incorporated CPU, memory, I/O into a system. Users time shared and also shared memory. Now some are starting to virtualize various components to improve operating efficiency. New systems are being configured with virtual machines, networks, and storage and accessed by multiple tenants. These different configurations raise various issues but currently there are no clear answers.

Once we establish the semantics of the backplane, we will see a new stack. This will include service networks, services, customer/core services, virtualization, all in some hardware plant. Building these big computers requires designers to address the issues of latency and bandwidth as well as synchronization waits. Fortunately, bandwidth within large systems is increasing, and the various link types are converging into standard protocols.

These exaflop systems are likely to run on 40 Gb Ethernet links which will be comprised of four 10 Gb links. This is about the same speed as for PCIe 3.0 links. The challenge here is that the time of flight through 10 m of optical fiber is about 50 ns. The serialization delay for 256 bytes at 40 GB per second is also 50 ns. Therefore, next-generation systems will all become optically switched multi-wavelength networks and backplanes.

Wavelength becomes an interesting characteristic. Interconnecting massive systems quickly runs into the limits of light speed. A 1 ns delay is approximately equal to 1 foot of copper wire. Creating highly dense machines can minimize the speed of light latency, because density goes up with the volume, a cubic factor, compared to I/O increasing with surface area, a square factor.

To enable these massive systems, we must get away from commodity thinking and consider the difference between computers and computing. The scale of the systems will drive brutal efficiency. Therefore, designers must consider these four principles: efficient is good enough; symmetric multiprocessing is too expensive; failures are frequent; and these changes are not going to happen by just changing the ISA.

Efficiency can be seen in areas like bandwidth balancing. An aggressive memory network ratio might be 20:1. Congestion occurs as a function of unstructured traffic and drop packets are latency singulants. SMP can never be big enough because of the need for high bandwidth and low latency. Multi-core is just another name for SMP. The problem here is that non-uniformity is a bug and not a feature in all architectures. The basic hurdle is the cost of changing nodes.

Failure must become acceptable. Local memory must have ECC and chip kill, because the compute pipelines are getting better. Changing the ISA to use more cores is not a complete solution either. Instead, changing the base CPU and ISA to become network aware is necessary. This illustrates another computer science failure, but a bug is the same as binary compatibility. An ISA includes NMU, bios, I/O, other hardware and also needs APIs for software extensibility.

Parallel distributed systems need to verify binary contracts. A compute node includes an OS and well defined memory, virtual and protected spaces plus other intermediaries. The move to computer virtualization blurs the division between hardware and software. Hypervisor’s insulate guests from the hardware and perform dynamic binary translations.

So the definition of the next systems will have a network integrated into the ISA. These systems will provide user level messaging in a fully virtualized and protected environment with robust support for global address spaces. The systems will support transactions and have some type of persistence mechanisms as well as clear exception mechanisms to keep sessions alive and working. Nonvolatile memories will store persistent objects.

So the future history of computers will look like this. Scalar, to vector, to MPP, to SMP, to clusters, to cloud, to epoch. Supercomputers will become commodity hardware and operating systems accessible to anyone.