Big jump
Surely nanoservers will happen before picoservers?
Server chip juggernaut Intel is hosting its Forum Beijing summit with partners and customers this week, and Diane Bryant, the new general manager of the company's Data Center and Connected Systems group, talked a little bit about some forthcoming Xeon and Atom processors for microservers. It took Intel a little while to warm …
I really don't know why little things like this haven't happened in a big way already. Just how much juice does it take to maintain a web session?
Add in the ability to switch bits of itself off and on to match demand, have a few powerful-per-core nodes to handle common tasks that might need the oomph, and that's a lean, mean, green-credentialled web serving machine right there.
"I really don't know why little things like this haven't happened in a big way already. Just how much juice does it take to maintain a web session?"
Well, there are a few things that might count against it... as a project, or small scale idea, it has some appeal (even if only to a hardware hacker) and is very close to how current blade servers can be tricked out. However at any reasonable scale this would most likely be a maintenance nightmare - the MTBF of consumer level stuff, like smartphones, running 24/7 would probably mean you'd spend a rather large amount of time just swapping the boards out. Pulling data off the disks also takes power, and even holding it in RAM takes a surprising amount, so the savings there are probably small to zero (the requirements for robust storage would still be the same).
There are some reasonable compromises (using current tech), such as the dense compute that the likes of SeaMicro make, and the T1/T2/T3 Sun machines - where the network, I/O and management infrastructure is rock solid and you have a very heavily threaded processor (with a corresponding low power requirement per core *). That said, there is a very definite move to combine these types of machines with even lower power cores, hence all the ARM server talk over the last few years, so you may get a version of your box quite soon - interesting times ahead.
* A dual-CPU T3-2 machine with 256 threads and 16 DIMMs running 100% load would draw about 600W, of which about 80-100W is taken by the memory. That works out at about 2W per thread, so from a purely instruction processing point of view we're already near some of the figures in the article.
...as I was thinking of things like the 42U rack unit with hundreds of ARMs in it that was featured here once.
Although a literal cluster made from Smartphone PCBs would be amusing, practical or not. I'm just thinking along the lines that something like web serving can potentially be quite massively parallelised. Given that some SoC or SoB arrangements can end up taking just a couple of watts each or less, that's matching or beating the performance per thread of the T3-2 machine, even assuming the boards are only capable of maintaining one session each. Then you can dial the thing right down so only one board is operational, so 2W (potentially) to keep an entire cluster ticking over in an idle state. Good, no?
Okay, not suitable for all loads, but I can see why the drive toward tiny servers with lots of cores is happening.
Aren't the Atoms the crippled chips that Intel developed for the netbooks that are no longer selling? They were also supposed to make their way into set-top boxes and phones IIRC. Seems that isn't happening so they're being repurposed as server chips. Crippled, of course, so as not to cannibalise the higher margin real server chips. In that comparison 6W is pretty impressive but what workload will they be handling? They might be able to virtualise but how many VMs will they comfortably handle? They might be 64-bit but somehow I don't see them being targeted for data warehousing.
We still have a while to go before the ARM-based servers arrive in force where OPS per Watt comparisons will be able to made. Yes, we know they won't be fully 64-bit but will it really matter for these smaller servers?