The biggest take-away for me is this gives you an oppertunity to (somewhat) rebalance the performance equation of the server.<p>Currently, your typical dual-socket, quad/six core CPU, is quite top heavy: LOTS of CPU performance, only a satisfactory amount of RAM bandwidth (and in certain cases not limited by bandwidth but by the DRAM command rate), along with piddling disk and network performance.<p>You could create a dual or quad-core ARM Cortex A-9 chip with a dual-channel memory controller (more memory bandwidth/ops per degree of CPU performance), dual or single gigE (more relative bandwidth per node), and finish it with a single SSD (lots of IOPs). Since these nodes are fairly affordable (probably $500) and relatively low power (probably 10-15 watts, you're left with nodes that 1) offer 1/3rd the performance but are 1/5th the price and use 1/20th the power 2) can be scaled cheaply and predictability, particularly as it pertains to laying out your datacenter power and cooling infrastructure.<p>Also, for people who don't have complete control of their data center (they're renting rack space), it eliminates a significant amount of headache finagling your datacenter to provide enough amperage to your rack, and eliminates surprise deficiencies in their cooling infrastructure.... lessons I've learned the hard way. Micro-servers let you scale at a much more granular level so that you aren't rolling out several kilowatts of servers at a time, and let you avoid carrying excess capacity as you wait to grow-into it (only to scramble to scale after you're full-up again)
This is so not my area of expertise. However, I wonder, wouldn't more small servers imply more hardware failures? Exchanging one HD in a big server might be less work than exchanging 10 HDs in small servers. Of course the HD in the big server might take down more "applications", but I think that is taken care of in other ways (hot swappable HDs, redundance?).<p>Overall I wonder if maybe a new OS for the cloud is called for? It seems inefficient to have seperate VMs running full OS for every tiny application. Maybe in the future not only storage will be a service (like S3), but also CPUs and RAM could be plugged together at will? Like there wouldn't be lots of small or big server instances, there would be farms of CPUs, farms of RAM, farms of storage, that could be combined at will. Maybe networking would be too much of a bottleneck, though :-/
<a href="http://www.cs.cmu.edu/~fawnproj/" rel="nofollow">http://www.cs.cmu.edu/~fawnproj/</a><p>tl;dr IO is the bottle-neck, not the CPU. So slow CPUs on lots of nodes gives you more cost-effective (in terms of power costs) key-value lookups.<p>Of course, this makes layer-7 load balancing a huge problem, which is my own little project in this world. More on that later :)
I'm hoping that these will shortly a) exist and b) be produced in enough volume that they unit cost will come down to <$100 levels. I'd like to buy a small, low-power cluster for hobbyist use but the current low-water mark is about $200 plus memory (which is limited to 2G). Hopefully there would be some standard on DC-DC distribution but that would just be gravy.
Traditionally, Intel and its partners have prevented the microserver market from taking off by limiting their Atom (and other low power) platforms to small amounts of RAM, usually 2-4GB.<p>It sounds like they're finally willing to open this up and allow a decent amount of ram (8GB and beyond) on low-power, tiny motherboards.
I need to do more research, but you could build one of these for about $200-$300 depending on volume. Assume that the storage would be an SSD connected a larger storage block.