For the benefit of those of us who don't know much about Atom, could you explain what the limitations are of using Atom for webserving?
I was surprised to discover that with a workload composed entirely of XML parsing using Python+lxml, an OVH 1.8ghz Atom N2800 beat hands down my mid-2010 2.4ghz Macbook Pro. This isn't a crazy result either. For instance, while the Macbook is equipped with dual cores, each core supports only one hardware thread, and thus is blocked while waiting to complete a load from RAM (a common occurrence when dealing with giant cache-unfriendly DOM trees that don't benefit from the Core 2's much increased L2).
In contrast the N2800 has 2 cores with 2 hardware threads each (hyperthreading), so each core can progress while its buddy thread is stalled. In my particular workload this was enough to beat by a significant margin a much faster, expensive and power-hungry processor.
Core2s have significantly better memory bandwidth, because they're dual-channel. They stream memory significantly more efficiently than N2800 Atoms. When a Core2 Duo has a memory stall, it can start executing other instructions out of order to compensate. Typically, out-of-order cores seem to do more for efficiency than hyperthreading.
If Apple screwed up by only giving a single DIMM (it happens on lower-end hardware, but I'd be pissed if I learned that an expensive Macbook were improperly configured memory-wise)... maybe the Atom would have a chance. But a properly-configured dual-channel RAM Core2 Duo, as old as it is, still would stream data faster from RAM than an Atom could. Almost twice as fast in fact.
If the N2800 does perform better on this workload, I'd be definitely interested in seeing the exact numbers... if you don't mind.
I actually do own a Hyperthreaded Atom (N570) and Core2 Duo (some old crap desktop version) myself. So... I'd be able to repeat the experiment :-p
In my case I had 1 software thread per thread, I'm not sure if over-subscription would make any difference. Also, out of ordering processing is only possible in the absence of data dependencies (I have no idea how to even go about measuring this). Also tree structures are no different to random walks from the hardware's perspective, i.e. the hardware has no useful predictive ability to prefetch data in this case.
The main point was that it's so cheap to test, there's little value in speculating about architecture specifics few are qualified to understand, assuming they're publicly documented in the first place
You're right in that its extremely easy to test, even to buy yourself an atom at home. Netbooks are regularly under $300, Clovertrail Atom Tablets are ~$400. Older Atom 330-based Netbooks are probably even cheaper. Its not like you're trying to test out the performance of a 64-core Quad Opteron or something.
And of course, the ability to just rent one for 3 months at a time from Kimsufi for less than $20 is always available.
The Atom is Intel's line of chips designed to compete against the iPad. It is still fully x86 compatible, its just slow. Roughly... the speed of a Pentium 4 or so.
Current generation Atoms are 32-bit only, which means 4GB maximum RAM (not an issue for smaller webservers with only 2GB).
Atoms tend to be best for I/O limited tasks. If you do video streaming for instance, the vast majority of your CPU power is going to be "wasted" on waiting for the Hard Drives and the Network. So the Atom makes a good, cheap CPU for that kind of task.
For tasks with tons of computations, (ie: game servers), the Atom is woefully inadequate.
You can purchase an Atom computer for ~$165: http://www.superbiiz.com/detail.php?name=MB-ZID60-U. So play around with the performance yourself, and see if its enough.
The Atom came out in 2008, two years before the iPad, and come from a lineage started in 2003. It was designed for the now dead netbook category that required low power consumption.
64-bit versions have been available since 2010, both versions used by OVH are x86-64. There are also server-specific versions with support for ECC memory, though they doesn't seem to be using them.
Regardless of the 'ordering', the Atom line of processors are designed for the <10W form factor. True, Atoms existed before the iPad, but they are in fact designed to compete in that power-range. The analogy works very well: the iPad 4th generation A6x processor is just slightly slower than the newest Clovertrail Atoms. The older Atoms (like the 330) probably are a bit slower than an iPad.
http://www.anandtech.com/show/6522/the-clover-trail-atom-z27...
So yes, perhaps I misspoke earlier. Nonetheless, I stand behind the analogy. Intel Atoms are Intel's version of the iPad A6x. They use roughly the same power and give roughly the same performance... but Atoms give you full x86 compatibility.
Perhaps a more appropriate way to say it... is that the Atom is designed to be a competitor to low-power ARM chips in general. Intel is making strides in making extremely slow, but power efficient chips. And in Performance/Watt, they're roughly on the same scale.
So Intel wanted a fairly power efficient chip, but they also wanted to ensure that it didn't cannibalize their own sales, so they intentionally crippled it from a performance perspective, not least by always building it on the last or second-last process. This is an aspect of the Atom that is missed by so many, sure that Intel was caught with their pants down by ARM: Intel's biggest fear wasn't ARM, but that their pricey high-end CPUs would get replaced by their low cost variants. They still fight with this paranoia to this day.
The Atom was neither inspired by the iPad (obviously given that it far preceded it) or even ARM. At best you could say Intel had some concern about Transmeta, leading to some of the early Atom work.
This isn't a minor error of "ordering". These were critical mistakes in your retelling of history.
I have admitted my mistake with an addendum, although I cannot edit my earlier post anymore. What more do you want?
This is incorrect. The Atom CPU was initially meant for smartphones, but the first models way overshooted the available power envelope. Netbooks were created by Asus when Intel offered a lot of cheap, slow chips for sale that were good for little else, and they created a new, shortlived segment. No-one at Intel thought about netbooks when designing Atom.
The first Atom processors actually had extremely competitive power characteristics. The only supporting chipset, however, was a) a power pig, because Intel just didn't bother despite it being a much easier task than making a processor (I have a little file server running a dual-core Atom, the processor being air cooled while the chipset has a fan. As a humorous aside, altogether that micro-PC uses about the same power as my new i7 Mac Mini under standard loads), b) completely built for traditional PC uses. Do you need IDE on your smartphone, for instance? How about an A20 line?
The "glut of inventory" notion is not realistic. It doesn't work that way.
Intel had no intentions in the smartphone industry at the time. If they did, not only would they have actually made a decent chipset (they happen to be the most advanced chipmaker in the world, with the world's best fabs), they would have made a prototype (which is what all manufacturers do when they try to enter a market - a proof of concept). There were absolutely none until later Atom variants.
People forget it now, but it was initiatives like One Laptop Per Child that initiated the netbook craze -- that program talking about ultra-inexpensive laptops, leading to a lot of people saying "You know...I too would like an inexpensive laptop that I could just bang around and leave in the car and..." (Slashdot was full of people trying to repurpose such devices), and with that an industry was born.
They never "really" tried to compete in the smartphone market (which is ~2W), the Atom was instead always targeted at ~10W. (IIRC, Intel's real focus at that time were Consumer Ultra-Low Voltage, CULV processors at the 18W envelope. Today... known as Ultrabooks). However, as Smartphones became the next hot thing, it is now obvious that Intel has to scale the Atom down even further. And unfortunately, the netbook market imploded.
Its not all wasted however, IMO, netbooks "became" tablets. From a CPU perspective, there is little difference between Netbooks and Tablets. You want low-power consumption between 5W to 10W, and lower costs. In fact, as netbooks were "dying out", they slowly became tablets. See AsusTab Smart for example, it basically is a Netbook without a keyboard, but built with a touchscreen. (same price range as Netbooks of old, but just in tablet form).
Anyway, Intel hasn't forgotten the Smartphone market either. Thus the Atoms that worked closer to the 2W SoC benchmark... Medfield, Clovertrail, and soon to be Bay Trail Atoms. Clovertrail is winning a few benchmarks here and there, and Bay Trail will be built on Intel's latest process for once. So it looks like Intel is finally taking the smartphone market seriously. But only the last generation or so are actual attempts at a Smartphone chip... and even then, there isn't a single Atom out there that is built on Intel's latest 22nm process. That will only come from the future Bay Trail Atoms.
To add some numbers:
The most popular first generation Atom, the N270, had a TDP of 2.5 W. The accompanying chipset and I/O controllers however had a combined TDP of 9.3 W.
Source: http://en.wikipedia.org/wiki/Bonnell_%28microarchitecture%29...
Some lines of Atoms since mid-2010 or so have supported x86_64. The Sxxxx "server" atoms released at the end of 2012 are all 64-bit-enabled, for example, although they can still only take 8GB of RAM: http://ark.intel.com/products/series/71265
I agree that this is mainly interesting if you're I/O-limited. Most benchmarks show a "regular" recent-gen Intel CPU performing around 5x-10x the Atom, so if you're CPU-limited, even a low-end VPS that gives you an average of 1/4 share of a core will probably outperform sole tenancy on an Atom core. But a high-contention VPS in most cases will give you much worse I/O performance. Also, the storage space and bandwidth quota here are much better than with a comparably priced VPS.
The specific issue with Go was significantly mitigated in Go 1.1 by making much more of the GC precise rather than conservative. Here's a comment from about a month ago on the current state: http://code.google.com/p/go/issues/detail?id=909#c59
Note that I've seen people running gnome on a netbook.
Still, I see this, and I think more about: "backups", "configuration management", "shell access", "VPN to other more powerful dedicated servers in OVH", "personal storage", "scheduled tasks", "monitoring systems", "status pages", etc...
The price is nothing, I still ask myself if it's rentable because "it's an Atom" and changes the power bill, or this is more a unprofitable product to attract you to other products.
Lets say 3€ * 12m = 36€ year, It takes some years to recover the hardware costs (without count electricity bills, company salaries, connectivity, taxes, etc).
pretty sure it's the latter. Even if there is minimal human effort on part of OVH involved that probably leads to unprofitability. For example a RAM stick breaks and has to be switched out, I guess it takes a systems engineer in the NOC at least 30 minutes to locate the server, disconnect it, replace the stick, mount the server again, make sure it is running fine. Assuming that employee get's paid 20 GBP, that is a 10 pound cost plus replacement RAM which is...I dont know...let's say 20 GBP, recouping that kind of money would take about a year (I know very dirty napkin calculations).
I bet they have a handful of these gimmick servers and they are going to sell out very soon. Comparable to the Raspberry Pi colocation edis offers ( http://www.edis.at/en/server/colocation/austria/raspberrypi/ ) try to order and you will find that they are sold out.
These servers run without cases, using water cooling. Hardware failure automatically notifies the NOC with the location of the failed hardware and replacement parts are already waiting on the desk.
>handful of these gimmick servers
OVH is one of the largest ISPs in the world, and is well known for disruptive pricing and services, such as 100/100 Mbps broadband to the home for $20/month.
Edit: Order placed, eagerly awaiting my Isgenug (that is what they are called for the German market) :)
As such their suitability depends greatly on your particular application.
They aren't so good for high concurrency for instance, but if you are using a mainly event driven server where concurrency isn't a massive issue you'll not notice that limitation.
At this price, the most cost effective option is going to be to rant one for a month or few and give it a try. In fact at that price if the machine works for you it is an absolute bargain, you could pay more for a much smaller VPS where you are competing with others for I/O bandwidth. I have a machine at OVH currently for a few bits and bobs - for my use pattern I might be better off cancelling it and getting a few of these instead for the same price.
As such, it can be quite surprisingly slow, managing to limit a lot of loads that most devs take for granted these days.