New on LowEndTalk? Please Register and read our Community Rules.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Comments
>
I would use nginx anyway.
Yes, but if your code isn't tuned for ARM, it will take longer enough to execute vs x86 that there might not even be significant power savings. If it takes 3x longer to execute at 5W, that's 15W. If it takes longer by 18x, that is effective 90W, more than many Phenoms that will give you the answer faster.
The x86 will also sit idle a lot with the same workload. That's the biggest waste of power.
uh, yea, sure, if you use a Q6600 or a Core2Duo - New Intel (and probably AMD) CPUs can scale down to almost zero when not in use/required.
www.hardwaresecrets.com/everything-you-need-to-know-about-the-cpu-c-states-power-saving-modes/
You should try to get a board and try for yourself
@bene_online When are you going to fix SIS?
Had random hang issues with a new system that pestered me until I got the idea of disabling something C-state related in BIOS. But it's a good idea, hoping the quality of implementation goes up on that.
How much do you think they are paying per board, when do you think the rate of return hits?
The boards are built "by them", they probably ordered a very large amount, the CPU & RAM is fairly cheap.
Id say what cost more is the bandwidth and rack space than the actual hardware.
Just think the cost of a Raspberry Pi, and that's with a middle company/distribution network making some %
They have the bandwidth, being part of one of the biggest french ISP. They need the rack space. And the power; but those babies don't eat too much...
Yea, Illiad uses a lot of incoming BW for their endusers - They likely have 100G+ unused inbound BW.
Rack space outside of Paris is pretty cheap as well, as is power if you have a large corp as buyer or a large baseload (or in online case BOTH).
They probably ordered 1000-10k boards from China at 40-50$ each (SoC 20$, RAM 15$, everything else 10$), no problem.
Never meant that its cost them the bomb, and I know they have plenty of spare capacity down to the fact of who they are, so that's the last thing they would worry about.
But I meant in the overall scheme of things, the hardware cost is probably minimal compared to the ongoing costs if they broke it down.
I agree with you.
The only downside is if they can't rent them enough to get back into the initial cost. I guess the future will tell what people consider they can do with those little boxes!
I agree, I can imagine the turnover of them will be quite high when people realize limitations and cancel after a month, I guess with the automation they are aiming on the churn of users to pay off the hardware in the long run.
Actually, I just have to say, these little boxes are pretty beast, you don't give them enough credit. Also, I wanted to mention for those who failed to understand this, your network bandwidth will be effects by your storage IO. These servers are all using network storage, which means that all of your IO is done over the network. You will notice if you start downloading a lot of data that your throughput drops off and this is because its using network IO in the background to make writes to the storage volume. Other than this shortfall, it is quite easy to run a lot of applications on these servers.
They would be GREAT for settings up a cluster or distributed setup, ESPECIALLY with the floating ip addresses. You could actually use a heart beat monitor and their API and have it switch your ip over to another server if one was to crash. You could easily use 4-5 of them in a cluster and have a more productive outcome then slapping it (a site/service) all on a single dedicated server. There are a lot of good options here with this setup, you just need to think outside the box you have been taught to sit in.
I can compile and update the server just as fast as some of my best VPS servers, and to boot, expense for the server is a fraction of the cost. With 4 full 1.3Ghz cores to abuse to your self, you can't really beat that for the price.
my 2 cents.
Cheers!
Actually in this case not as much, since downloading is incoming, whereas storage writes are outgoing, and you can use both at the same time.
Otherwise yeah it's something to keep in mind, but not a terrible problem, as if you utilize the specified 200 Mbit Internet bandwidth, you still get plenty left (800 Mbit, or about 100 MB/sec) for disk I/O.
I have one and was trying to push a bunch of data to it, it got bogged down trying to write out and my transfers droped from 40M/sec to around 18-20M/sec. Once IO caught up it would again bounce back up to 40-50M/sec. That's just been my experience. Also, keep in mind that most gigabit nics in these arm boards are tied to CPU (One of the reasons why BananaPi A20 can't get above 50M/sec as that is the max one core can provide) and the irq can only be assigned to one core, so your speeds can vastly be effected by how busy the core the network irq is tied to is. One way I have helped boost speed 'some' is to assign the NIC to CPU 2 instead of CPU 1 where work by default is started.
16: 15008680 15237937 14815187 14887786 armada_370_xp_irq 5 Level armada_370_xp_per_cpu_tick
18: 0 0 0 0 armada_370_xp_irq 50 Level d0010300.rtc
19: 394 0 0 0 armada_370_xp_irq 41 Level serial
25: 15404344 25990 0 0 armada_370_xp_irq 8 Level mvneta
>
>
>
My little tip for the day
Cheers!
Anyone noticing slow http downloads? Is each http connection limited to around 100-150KB/s? If I use a download manager, it maxes out my connection, if I just use the browser it just hovers around those numbers.
Sure! Their DC is in Vitry Sur Seine, near Paris.
Shenzhen can do everything.
I assume you're compiling on the box itself? Compiling to ARM as a target is not the same benchmark as compiling to x86 as a target, due to completely different optimization algorithms, or for those algorithms that are the same, due to being run on completely different input data with completely different constraints.
I do get that these things have uses, especially if you're not running a lot of complex stuff server-side and that's why I was interested if anyone knew of a similar service in the U.S. (in another thread).
REALLY NOW, a quad gigabit SoC designed specifically for servers would have its NICs tied to CPUs? Sure why not, raspberry pi has, so this one must have too!
You can use NAT with iptables to get internet access to the 2nd server over the private network.
Yep, but then you'd need two servers. There is no option for a NAT/shared IP. That would be great for applications that don't need a dedicated IP!
Yes and it's an awesome option for a dedicated database backend server.
Stateful NAT (many private IPs to a few public IPs) at these tens of gigabits speeds is very expensive to deploy, requires either a top of the line DC-grade router, or a cluster of powerful x86 servers dedicated just for that task. I bet nobody is going to budget for this just so that you can save 1 EUR on IP. Now when IPv6 get deployed, you'll get IPv6-only servers, but I'd expect still no IPv4 NAT.
Thanks for your answer.
More IPv6 only offers would be great - and you could have a 4 to 6 reverse proxy easily set up somewhere
Actually, most of the issues with C-state come from bad/old power supplies that cannot supply the minimal amount of current needed.
http://www.bit-tech.net/news/hardware/2013/05/01/haswell-psu/1
Pretty sure I wasn't using an "old" power supply, but it could've been a "bad" one. Could be Linux's fault too, there are many chipsets where Linux needs a few kernel options passed to it to run smoothly, and it's very hard to find out which ones.
@rm_ I appreciate that you want to just state things like you know them, but how bout if your gonna shit on me and act like I don't know what I am talking about you show the documentation to back up what your saying? You say it is different and will have no effect, yet I get better throughput by assigning a different core, so either its a fluke or it does have an effect. I am glad you're participating and I know you have put some time into arm your self, but please if your gonna just say I'm full of shit show the documentation to back it so we can all learn something. I am not perfect, but nor are you so stop acting like it. I am more than happy to admit I'm wrong if you can demonstrate it.
Cheers!
got one
i love it.
Point is that Ethernet on these boards is much more efficient than on something designed for a TV set-top box or a media player (like the A20 in your Banana Pi), and throwing them all into the same heap by saying "these arm boards", is misleading at best. It's wrong to take some limitation you faced with the A20 and automatically assume this board must also have it. Sure pinning interrupts to a specific core helps optimize things, but that's nothing new, people do that even on x86 platforms.