Weekly edition	Kernel	Security	Distributions	Contact Us	Search
Archives	Calendar	Subscribe	Write for LWN	LWN.net FAQ	Sponsors

KS2010: Scalability

Posted Nov 3, 2010 22:55 UTC (Wed) by cesarb (subscriber, #6266)
Parent article: KS2010: Scalability

> He said that silicon technology is getting closer to the absolute physical limits, so we'll not have to worry about CPU scaling for a whole lot longer.

Wouldn't getting closer to the physical limits make things _worse_, because instead of just making a faster CPU, people start having to add more and more separate cores? That to me looks like needing even more scaling, not less. Even more if the latency between the cores increases because there are so many of them.

(Log in to post comments)

KS2010: Scalability

Posted Nov 4, 2010 2:34 UTC (Thu) by PaulMcKenney (subscriber, #9624) [Link]

You can argue this any numbers of ways:

As you say, if we really need exponentially increasing numbers of CPUs, then silicon limitations will require more sockets and thus more latency.
If the limits are economic, then most people will buy as large a system as makes economic sense for their workload. This has historically resulted in very many very cheap small systems and a very few very expensive large systems, effectively limiting system size more severely than the pure technical limits.
There is the question of whether the limits are good or bad, in other words, do you favor increased CPU counts on the one hand or existing software on the other?

KS2010: Scalability

Posted Nov 4, 2010 12:40 UTC (Thu) by dgm (subscriber, #49227) [Link]

Fear not. Costs increase witht the number of sockets a system can support, quickly reaching the point where it's not economically viable except for very specialized (and thus expensive) systems.

KS2010: Scalability

Posted Nov 4, 2010 12:45 UTC (Thu) by nix (subscriber, #2304) [Link]

Yes, costs increase with the number of sockets... which is why chip manufacturers are adding extra cores instead. Hence the problem.

KS2010: Scalability

Posted Nov 11, 2010 22:11 UTC (Thu) by jnareb (subscriber, #46500) [Link]

There is also speed of communication limit, which coupled together with miniaturization limits (which means that length of communication channel would grow with number of cores) makes increasing number of cores a moot point. Adding cores would not increase performance, except in rare cases where synchronization is not needed.