SeaMicro drops an atom bomb on the server industry

http://venturebeat.com/2010/06/13/seamicro-drops-an-atom-bomb-on-the-server-industry/

by timf • 16 years ago

260 109 16 years ago

stcredzero3• 16 years ago

The way I read this, they are achieving savings by virtually mux-ing (or de-muxing depending on viewpoint) much of everything that's not the CPU. Is this optimized to make the support of virtual servers with relatively low throughput more efficient?

njharman• 16 years ago

I read it to be they created a chip that runs code written to emulate/virtualize most of the other (non-cpu) logic on a motherboard. so instead of 10 "io" chips and 10 "memory" chips most of which are idle they have 15 virt chips which may act as either "io" or "memory" depending on demand.

I guess (de)muxing is close to that? Multiplexing many "request" amongst fewer real(non-virutal) objects.

stcredzero3• 16 years ago

I read it to be they created a chip that runs code written to emulate/virtualize most of the other (non-cpu) logic on a motherboard.

That's exactly what I meant by "virtually (de)muxing everything but the CPU." I should have said, "CPU+chipset," though.

I wonder if their architecture is also (unintentionally) optimized for languages like Ruby/Python? I suspect that languages like those tend to have more CPU operations per IO operation. I wonder if anyone has researched this metric?

hga10• 16 years ago

Well, they're only muxing IO; each Atom CPU has a chipset CPU and it's own DRAM (up to 2GB (the Atom has a hard 4GB limit), no ECC since Intel only offers that for official server chips).

GFischer• 16 years ago

Interesting... I've never been in a really big datacenter, so I'd like to see some (hopefully non-biased) reviews from somebody that does work in those places.

Would this really work well for the intended market? There are lots of startups over here that plan on massively serving webpages - would something like this (only cheaper :) ) make you reconsider using whatever cloud services you're currently using?

Articles I found on Google:

http://gigaom.com/2010/06/13/seamicros-low-power-server-fina...

and the Wall Street Journal's take:

http://blogs.wsj.com/digits/2010/06/14/seamicro-tries-to-ret...

stcredzero3• 16 years ago

There are lots of startups over here that plan on massively serving webpages

Seamicro is using the same sort of lateral move that RISC made and that Transmeta attempted to apply: Could certain functions happen more efficiently if moved outside of the traditional 'box?'

Could a datacenter provide a low-latency infrastructure for front-end web-servers while reducing their power and other expenses? I think there's a good chance this is already being done.

lsc12• 16 years ago

what I see right now is that people are choosing fewer, higher speed, more expensive and more power hungry cores over more, slower, cheaper and more efficient cores. Anyone who uses Xeons instead of the new 8 or 12 core opterons has made this choice. Sure, the Xeons kick ass when it comes to single threaded performance, but as far as performance per watt or performance per dollar go, the new socket G34 opterons kick the shit out of the Xeons. For the cost of an 8 core Xeon, I can get a low power 8 core opteron, 32GiB ram, and a nice SuperServer to put it all in.

Andys• 16 years ago

Its a custom-designed niche product. It will be priced accordingly.

You can tell from their "dollar savings" chart that they plan to claw back the money you save in price of the machine!

p3ll0n• 16 years ago

James Hamilton, VP at Amazon and author of the popular blog "Perspectives" has some interesting insights into and great things to say about the work SeaMicro and other new startups are doing to revolutionize the server industry.

http://perspectives.mvdirona.com/2010/06/14/SeaMicroReleases...

ableal2• 16 years ago

(timf had submitted that one too, but it slipped down)

Another write-up with names and tech details here: http://www.theregister.co.uk/2010/06/14/seamicro_sm10000_ser...

hga10• 16 years ago

See http://news.ycombinator.com/item?id=1429625 for HN item on this excellent posting of his with a few comments, e.g. I note he notes this system has no ECC. And that Intel is pretty unique in that limitation.

astrodust• 16 years ago

I hope this shakes up the VPS hosting industry, too. I'd love to see what kind of Linode or EC2 style on-demand computing these might provide.

wenbert3• 16 years ago

Hopefully it will be cheaper. $20/month might sound very cheap but for some guys (like myself) living in 3rd world countries, it's almost next to impossible.

lsc12• 16 years ago

Hey, I have a question. If I wanted to advertise to /technical/ people in your area, where would I do it?

Note: I'm not really interested in business people; I want *NIX enthusiasts who are /not/ spammers, who can write English and are capable of running a Linux box that won't get compromised, and who don't need much help from me to run the thing.

lsb• 16 years ago

If you'd like really cheap web hosting, try http://nearlyfreespeech.net. Write something in PHP, and you can use a SQLite database for free (just put a high timeout on write locking) or a MySQL database for $0.01/day or something. They don't support FastCGI, so any other language would require starting up a new runtime for each web request, which is pretty much a non-starter if you can't afford $16/mo (which is the prepaid limit).

If you need money for a project to get kickstarted, try Kiva or Kickstarter.

On the plus site, if 5 of you can fit onto the EC2 small instance, it's only $6/mo, if you do a spot instance with a high maximum.

noahc• 16 years ago

ASO has $25/year web hosting that is solid.

lsc12• 16 years ago

eventually, maybe; but right now there just isn't very much price pressure in the VPS industry.

Personally, I won't touch any hardware until I can get it from more than one vendor (usually, until you get to that point it's too expensive. Their may not be much price pressure in the industry as a whole, but in my niche, there certainly is.)

bnoordhuis2• 16 years ago

  right now there just isn't very much price pressure in the VPS industry

Isn't that the point? Right now, there isn't much of a margin left to undercut competitors with unless you scale up to Amazon or Rackspace size. Developments like this could change that again.

InclinedPlane• 16 years ago

Indeed. It's a rung on the ladder to times when what fills a rack or a server room today will fit into a 1U tomorrow. There will come a time in the not too distant future when the ability to host a site in which quite literally a majority of the world's population accesses multiple times per day is simply a matter of setting up a relatively run-of-the-mill inexpensive VPS. The implications such technology will have on business, society, culture, etc. are left as an exercise for the reader.

Super745• 16 years ago

Big surprise, Dell, Hewlett-Packard and IBM not innovating? I can't believe it...

Just look at the computers they are selling. most still offers 512M of RAM as a default. What can you run on 512? Nothing like having to upgrade the day you receive your product. Consumers by and large need the guidance of manufacturers to make the right HW choices and the manufacturers just want the cash. They will suck each segment dry until the market forces them to make the changes. That's progress?

With all the advances in multi-monitor add-on SW or the fact that Windows and Mac OS have been supporting multi-monitor control for years, try finding hardware already fitted with multiple display connections. In the end, the user is forced to customize their own equipment. And yes, I know most HN readers do this, but I am talking about the general public.

Great to see the little guys are still fighting. The big ones really don't give a damn.

rbanffy• 16 years ago

> What the F* can you run on 512?

A lot, really. I have rendered DVD quality movies, handled product composition databases and processed millions of orders on machines with less than 512 megs of memory.

Not many companies (or people) have enough data or volume to stress a machine with 512 megs of RAM. I agree it seems a pittance - my netbook has more than that - but if I break up memory usage by application here, Firefox takes most of the RAM, then Gwibber, then Rhythmbox. None of them would be running on a server.

It would be interesting to write down a list of what you can't do with 512M.

Super745• 16 years ago

Well, you can't effectively run Mac OS X or W7 on 512 to start, the most used operating systems in the world. Even Ubuntu needs an upgraded graphics card and RAM to utilize all of the bells and whistles.

As I said before, I was referring to the general public. There will always be hackers or high-end users that can make miracles.

wingo2• 16 years ago

Interesting article, but a terrible title! Atom bombs, really?

jbarham• 16 years ago

I think it's supposed to be a pun...

wingo2• 16 years ago

Oh dear, I did not realize that. Internets one, wingo zero...

mattmcknight2• 16 years ago

It would have been better to see a comparison with SGI (ex-Rackable) CloudRack systems, which have a bit of an inbetween approach, using Xeons, but at least nominally seem to pack in more cores in the same enclosure size. One of their power tricks, in addition to pulling back DC converters further from the computers, is to allow things to run hot, resulting in savings on cooling costs.

Scott_MacGregor2• 16 years ago

Too bad the electrical signal on the board cannot be altered in some way to allow less heat production for a given amount of processing. That would be a win-win situation all around. The heat seems like wasted energy, perhaps inefficiency that could be done away with by thinking outside the box.

Tamerlin2• 16 years ago

The SGI machines are geared pretty heavily toward supercomputing applications. As such they're tailored for applications that need to share data between processes efficiently, and that also need to do a lot of computing.

The SeaMicro server probably would look pretty lousy when compared on SGI's turf, since Atom processors aren't designed to be performance-competitive with the current desktop and server processors.

What would make more sense is a comparison a datacenter for a company like Disney or Google or Amazon. I suspect that even though SeaMicro's rig would look pretty bad for heavy number-crunching apps, it would most likely do very well in web and database. Those apps don't need to share a lot of data between threads/processes, and they're generally I/O bound rather than compute bound. With proper caching setups, they're mostly network bound, rather than disk bound also.

mattmcknight2• 16 years ago

I was thinking of the Rackable stuff, not the supercomputing stuff. They actually acquired SGI, but took the SGI name. http://www.sgi.com/company_info/newsroom/press_releases/2009...

pornel• 16 years ago

Their tech overview PDF has less layman fluff than the article: http://dev.seamicro.com/sites/default/files/SeaMicroTechOver...

hackermom• 16 years ago

An ARM-powered variant would sit on top of this Atom-powered machine, with quite possibly the same or higher numbers of efficiency when looking at performance-per-watt.

The article quotes the interviewee saying that ARMs can be used, but not clear enough to determine if they actually have ARM versions of these machines, which I'd be very interested in seeing 100k specint_rate figures for.

pjscott5• 16 years ago

Most of their technology is not related to the processors themselves, but to the rest of the computer. They've got a fancy high-bandwidth, low-latency interconnect between the chips. They multiplex access to I/O hardware, which I'm sure takes advantage of that nice interconnect. They have hardware-accelerated network load balancing which tries to keep the CPUs at maximally-efficient levels of utilization, letting the spare CPUs sleep, and avoiding malfunctioning CPUs. I would love to see this with ARM Cortex-A9 processors, and I think that's actually very likely if this company succeeds.

GrandMasterBirt• 16 years ago

I am completely baffled that Intel did not do this. I guess they figured ATOM is just a crappy little side-project processor, not one of the big boys.

megaman821• 16 years ago

They basically did it's called Larrabee and it is now being transformed into a server product.

pjscott5• 16 years ago

It looks to me like Larrabee's successor is is going after the high-performance computing market, not the server market. Why else would they devote so much die area to having loads of wide floating-point SIMD units?

ippisl2• 16 years ago

Some see the move to many-core ATOM servers , as a move to put financial pressure on intel(since ATOMs are cheap).

And this move helps the future of ARM based servers , since ARM is more similar to ATOM than to XEON.

hga10• 16 years ago

There's going to be limits to the amount of pressure as long as Intel doesn't offer ECC for any Atom variety microarchitecture. One of the articles I read said that they thought about using an AMD chip but it wasn't going to ship soon enough. I can see them hearing from their customers (would be and actual) that they'd buy more if they offered ECC and SeaMicro offering an AMD based box in the future.

I wonder how Atom specific the ASIC is....

Scott_MacGregor2• 16 years ago

Great news! Hardware innovation typically means new software opportunities. If this is turns out to be a generally accepted workhorse server design and not just a hotrod box I wonder who will be the first in here to develop a profitable software product for it.

Dell products have features and services that make them enterprise friendly, they are more than just hardware to the customer. So trying to compete with Dell head on might not be the company's best strategy at first. Perhaps following the strategy EMC used with CLARiiON to sell through DELL would be more of a money maker for the company.

patrickgzill• 16 years ago

Interesting in terms of technology, but of no interest to me as someone who does colocation and web servers for my clients, especially since they almost all use traditional RDBMSes like MySQL and Postgres.

The Atom is too underpowered and too RAM limited for individual systems - you would do better in most cases with a 2x quadcore setup and 32-64GB RAM combined with OpenVZ or Solaris Zones. Lack of ECC = automatic disqualification for me as well.

For a company that is doing a lot of web serving a la Facebook or eBay I can definitely see the appeal. In such larger cases, power usage dwarfs many other considerations.

arvinjoar• 16 years ago

I would rethink that if I were you. What if you started to provide virtual private servers on one of these boxes instead of offering colocation? I think this innovation makes VPS solutions even more efficient than they are today.

lsc12• 16 years ago

ECC is a huge thing in that case.

The thing about being a VPS provider is that if your hardware chokes, your customers notice, and they get pissed. Sure, if you are running some kind of web farm you set it up so that things keep running without interruption if you lose a server. But if you provide VPS hosting and you lose a server? People notice, and they will be pissed. If you are a VPS provider and you loose customer data? they will parade your severed head around town on the end of a pike.

Amazon has successfully tackled this problem by training their customers that they can shoot a server at any time. This is good for them, but it's not how VPSs have traditionally operated. customers expect the thing to stay up and for their data to stay safe.

lsc12• 16 years ago

hm. the price data would show it's not more efficient in terms of capital. $140K, right? before disk? for 1024GiB ram and 512 CPUs. with my current setup, I end up paying about $1900 for 32GiB ram and 8 (/much/ more powerful) CPUs before disk. So, I'd need 32 of those puppies, at a much more reasonable $60,800 - it would give me half as many CPUs, but, uh, from my experience a atom core is less than half a opteron core.

how much power would that eat? in my setup, we'd be talking about 3840 watts, each one of those single socket 8 core boxes eats about one amp of 120v, give or take. they say 4w per server, so that'd be 2048 watts for their solution; so yeah, it will take quite a while to make up the price difference.

Now there are some reasons why having your own physical server is better than having a virtual slice, but many of those reasons have to do with I/O, and as these boxes don't come with local storage... so at least for my workloads, they still can't beat the nice, standard opteron G34 systems I throw together in my workshop.

surlyadopter• 16 years ago

With that many processors crammed into such a small case isn't heat dissipation a problem? Or are the Atom chips used not as power hungry as a standard commercial cpu?

flogic• 16 years ago

Atoms are low on power usage. Hence their popularity for net books.

jonknee• 16 years ago

"Atom chips used not as power hungry as a standard commercial cpu" is quite literally the entire point of this startup/product and the article. You had to quit reading before the second sentence to miss this:

> The startup is announcing today it has created a server with 512 Intel Atom chips that gets supercomputer performance but uses 75 percent less power and space than current servers.

c00p3r4• 16 years ago

There are no information about the chipset, memory specs and extension slots in particular.

It could be a huge bottleneck between CPUs and RAM, because of concurrent access of so many CPUs to the memory while Atoms have a very small caches.

It means that real-world applications, like a multi-threaded services (especially JVM-based, or simply MySQL) could not be used efficiently.

I'm also very skeptical about hundreds of KVMs which is a the only working virtualization technology I know. =)

hga10• 16 years ago

Here's a lot more technical detail: http://www.anandtech.com/show/3768/seamicro-announces-sm1000...

Per CPU (8 per board):

  Atom Z530 (1.6GHz, Silverthorne)
  US15 chipset (Poulsbo)
  2GB DDR2

Per board:

  4 ASICs to virtualize IO
  2x PCIe x16 connectors

"Each hop takes 8 microseconds", so good enough for (most/all?) IO but not good enough for memory.

No ECC of course, it being a non-server Intel chip. Each single cored CPU has its own DIMM.

* * * * EDITED to clarify the layout per CPU and per board after c00p3r made it clear I wasn't being clear. * * * *

Reply

c00p3r4• 16 years ago

So, it's just a tight packing of a 512 netbook's boards in a one box. I thought it is a 512-core system. Oh, I'm such a moron. Nothing to see here.

Reply

wmf4• 16 years ago

It's a cluster, so each Atom has its own RAM.

Reply

c00p3r4• 16 years ago

Yeah, I missed the point. =)

Reply

zokier• 16 years ago

http://en.community.dell.com/dell-blogs/b/direct2dell/archiv...

Similar concept from Dell, from over a year ago. Although Fortuna seems bit more conventional, I'm not sure if it's a bad thing.

Reply

limist• 16 years ago

"People who are really serious about software should make their own hardware." - Alan Kay

This news is yet another data point that developers will need to hack concurrency sooner than later, as a core skill in one's professional repertoire. Off to learn Stackless PyPy, Clojure, Scala, etc...

Reply

c00p3r4• 16 years ago

"People who are really serious about software should make their own hardware." - Alan Kay

Where are Sun, SGI, Cray, etc now?

Reply

mbreese• 16 years ago

Is Apple just an anomaly? Also, Sun was bought partially so that Oracle could offer a complete software/hardware package, so I'm not sure your point stands. The others had great runs with great hardware/software... they just didn't survive to the next generation.

Reply

dharmatech• 16 years ago

The Cray Jaguar is the fastest on the TOP500.

I'm actually in Chippewa Falls at the moment. :-)

Reply

WilliamLP• 16 years ago

I think I can afford to wait until people start proving undeniably (through out-competing rivals) that Scala and Clojure have any pragmatic benefit in real multicore programming tasks. Anything that mentions Fibonacci numbers doesn't count, nor do easily parallelizable problems commonly handled in C/C++ (ray tracing, media encoding, web servers, etc.) It needs to be something not trivially broken up into independent parts.

Reply

drhodes• 16 years ago

While this isn't undeniable proof, it's interesting none the less, Simon Peyton Jones: Data Parallel Haskell.

http://www.youtube.com/watch?v=NWSZ4c9yqW8

He talks about how this scheme can parallelize non-trivial data structures, and how tedious book keeping can be avoided. I haven't used it yet in a project, so I can't speak of its efficacy - but worth a watch, imho.

Reply

hga10• 16 years ago

Don't forget that part of the claims are not that the resultant programs are more computing resource intensive but that they are more easily developed and debugged. Which is pretty important in this arena.

And while I don't know about Scala, Clojure is a little too young to obsess on preformance, each point release (1.0, 1.1, etc.) is delivering significant improvements.

Reply

xtacy• 16 years ago

This seems quite similar to the FAWN project at CMU. http://www.cs.cmu.edu/~fawnproj/ The idea is similar: if IO is the bottleneck, instead of scaling up IO, scale down the CPU power.

Reply

CountHackulus• 16 years ago

I do see a potential problem here. In the pictures, they show a bunch of Atom CPUs soldered directly to the board. That means dire things for service. Now, if a single CPU has a a flaw, you need to replace an entire board of CPUs.

Compare this to a standard blade setup, where you could just swap out CPUs, or even an IBM System Z where you could hotswap the CPU, and service doesn't look so great.

Reply

gfodor• 16 years ago

You're certainly jumping to a lot of conclusions here, perhaps the boards of CPUs are quite cheap?

Reply

hga10• 16 years ago

Or just ignore it if it doesn't take out the whole card. Atoms only come this way (solder mount), right?

You can hotswap each 8 CPU board.

Reply

wmf4• 16 years ago

To be fair, the unit of hotswap in System z is a book that contains ~32 cores; that's not so different from SeaMicro.

Reply

pjscott5• 16 years ago

The load-balancer monitors the health of the processors and can route around broken ones. Some amount of degradation is okay.

Reply

DEinspanjer2• 16 years ago

I worked with one of these prototype boxes. I was using it for something a bit outside their common use case: clustered ETL processing of log data. I was quite happy with the performance. In the workloads I had that needed lots of threads, I was able to use the box to spin up a lot of nodes and crunch through several hundred GB of log data very quickly. The machines were easy to work with since they felt like normal Linux nodes, and the interconnect fabric made inter-process communication very snappy.

Reply

runT1ME• 16 years ago

I'm still confused, is it a single multiprocessor machine, or is it a 'cluster' in the sense that each node has a single processor? (You used both threads and inter-process, but also implied multiple nodes...)

Reply

blhack2• 16 years ago

>I'm still confused, is it a single multiprocessor machine

I would love to watch that thing boot, haha. SO many penguins!

Reply

hga10• 16 years ago

It's a cluster, each node is a single core processor with a chipset chip, DIMM and 1/2 of an ASCI providing virtualized I/O (e.g. each core sees 4 virtual SATA drives) and an connection to a supercomputer style interconnect. Memory is not shared, each core is running it's own OS image.

Reply

DEinspanjer2• 16 years ago

Sorry I was rather vague there. Ultimately needed lots of parallelism. Smallest unit of parallelism in my ETL tool is threads. If you need more threads than one core or CPU can deliver, then you cluster out to multiple machines, and they communicate via sockets as needed.

Reply

pjscott5• 16 years ago

Performance per watt depends on the computer and the workload. It's not apparent from the SPECint benchmarks they show in the article, but chips like Atom are better than beefy server chips for some server workloads, and worse for others, when measuring performance per watt. Here's a paper from Berkeley's RAD lab which tries two Atom processors and a Xeon on several different server workloads, and compares their performance per watt:

http://www.sigops.org/sosp/sosp09/papers/hotpower_10_chun.pd...

The tl;dr version is that what you really want are hybrid datacenters, where you can assign various workloads to different types of machines, and use each machine type for what it's best suited to do.

Reply

ssp• 16 years ago

for a comparison of various types of chips.

Reply

pjscott5• 16 years ago

That chart compares theoretical peak GFLOPS per watt, which is an absolutely terrible metric for server workloads. As the paper I linked to showed, Atom is sometimes much better and sometimes much worse than large, fast server processors in performance per watt on real server workloads. Those charts don't take into account things like cache, or branch prediction, or multicore memory coherence.

The only way to get meaningful performance per watt data is to actually run real workloads on various processors.

Reply

ericd• 16 years ago

Yeah, there are still some single-threaded tasks that I want my servers to be able to tear through. For serving up HTTP requests, though, this sounds great.

Reply

rythie• 16 years ago

Does this really solve a problem anyone has though?

It seems we have an oversupply of CPU on modern boxes and an under supply of I/O speed (or space with SSDs) and memory.

Reply

jobeirne• 16 years ago

Where can I invest?

Reply

10ren• 16 years ago

I would think that ARM chips would use less power than Atom; but compatibility with existing software is a big selling point (as always). But this would provide a pressure towards ARM servers, and therefore ARM-compatible software.

OTOH, I get the impression that the bulk of the power savings are not from the CPU at all, but from virtualizing the other components. Therefore, the pressure towards ARM is much less.

And I suppose Intel wouldn't be supportive unless there was a compelling long-term reason to choose Atom.

Reply

sh1mmer• 16 years ago

Anyone care to suggest some ideas on a few things:

a) who would likely buy these (corporates, SMEs, startups)?

b) it seems that they have increased the risk of single point of failure (e.g. 1 PSU taking down 128 nodes) what's the mitigation strategy?

c) what would an architecture on a box like this look like? Should I just be thinking of it as a cheap set of VPS nodes?

d) People keep mentioning the kind of processing these chips are good for and not so good for. Can someone be explicit about good real world uses and bad ones?

Reply

hga10• 16 years ago

I suspect a lot if not most of their customers will be buying more than one, so that mitigates the PSU problem.

Reply

wmf4• 16 years ago

a) Web 2.0 sites.

b) It has four redundant power supplies, as do all systems of this size.

c) It's a Beowulf cluster.

d) See the FAWN, Amdahl blade, Gordon, and CEMS papers.

Reply

SeaMicro drops an atom bomb on the server industry

Related Stories