3 ms·
This is why I hate using AWS and similar virtual servers. I use them for quick dev stuff then get a non shared server. Massive datacenter scaling doesn't save t
by throwasehasdwi 9y ago
This is why I hate using AWS and similar virtual servers. I use them for quick dev stuff then get a non shared server. Massive datacenter scaling doesn't save that much money. The money in cloud is made from huge over-provisioning.
Most providers will artificially limit you below the published limits if you use "too much" CPU or internal network bandwidth. You think that 1Gbps link between your machines isn't shared? See what happens when you try to use it full tilt for a couple days.
- Elect2 9y agoAs far as I know, AWS and Google cloud instances do not limit your vCPU usage.(Exclude aws t1&t2 instance and Google micro&small instance) . And also they do not limit your bandwidth usage, they bill you instead.
- extra88 9y ago> The money in cloud is made from huge over-provisioning. Did you mean "overselling," selling more than your infrastructure could support if every customer used all you sold them?
- jsmthrowaway 9y agoI think you'll find that AWS doesn't do that outside of the vCPU credit-system nodes (which I think are just t1/t2). The resources you have rented from Amazon are your resources, especially as you get up to scale (when it really starts to matter in a big way, capitally speaking). I didn't make up that sentiment, by the way; an account representative said it to me verbatim. It seemed mostly true in practice, and it's a big driver of why it's tough for firms like Linode and DigitalOcean to compete in the same arena.
- i336_ 9y agoFWIW, I've compiled Chromium (which took a while because the build needed more RAM and swapped) and accidentally left random scripts erroneously chewing 100% CPU for hours (woops) on Contabo's $9 offering (6GB RAM, couple cores, 500GB disk, 100Mbps). Never heard a peep from Management™. Performance is mildly all over the shop - eg, some days packages might install in under a second, other days pacman stalls for 30s+ "resolving dependencies", and then a further 2-3+ minutes regenerating the manpage cache or whatever it's doing - but, you can peg both cores, and they don't seem to care. So there's that. It's in Germany, and ramp-up time to US servers takes a couple seconds (eg, pulling data from the US will take a moment or two to wind up to ~12MB/s). Similarly, latency is mildly high. A big plus: it's just QEMU and they give you access to the KVM VNC console. You can kill iptables when it goes rogue! You can disable SSH! Oh - a tidbit to remember: they do mention that if you use >60Mbps for 6 hours+ they kill your link, but I learned from some friends what that means in practice: your connection goes to 10Mbps for the rest of the month. Not a shill, just mentioning what I learned from using a friend's instance for a couple months. Would be interested to hear about any negative experiences people have had so I can file those away too. --- Another option: online.net have dediboxes based on Atom C2350s with 4GB RAM, 1TB and 1Gbps networking, for €8.99/mo (+€20.00 setup). So on the one hand you're dealing with this sadness: http://ark.intel.com/products/77977/Intel-Atom-Processor-C2350-1M-Cache-1_70-GHz?q=Intel%20Atom%C2%AE%20Processor%20C2350%20(1M%20Cache,%201.70%20GHz) http://ark.intel.com/products/77977/Intel-Atom-Processor-C23... But on the other hand it's a dedicated machine. The CPU has virtualization support (despite being 1.7GHz :D hahaha). That's gonna translate to 100% usage sooner or later... so... that may be a potential alternative. I haven't tested this option, it would be very interesting to pit it against a Contabo VM.