Commencement of commercial cloud operations

Selektel's promised cloud this summer, in which only consumed resources are paid, is finally ready.

In fact, we started a month ago, but decided not to arrange a habraeffect, but to receive customers a little bit. Thanks to this, we were able to carefully examine the difficulties encountered and solve problems in a very individual way. The strategy paid for itself, since during the “half-launch” there was one unpleasant typo in the template for one virtual machine, several minor flaws in the interface. At the same time, calmly, without pre-start confusion, additional functions of managing virtual machines were implemented.

However, enough about the workflow, I will talk about it later.

Selectel cloud is the ability to pay only for the resources used by the server. And the resources themselves are provided to the extent that is needed for fast and comfortable server operation. Automatically. Therefore, we do not have different tariff plans - we only have prices for resources, and the total cost is determined by the server load.
')
However, it is clear that even an idle server will consume resources (the kernel needs memory, the disk with the operating system must be stored somewhere). Anticipating questions: the monthly price of the most compact version of Debian with constant uptime and without significant load costs about 45 rubles per month.

And yes, you can turn off the machine, and the only thing you will pay for is storage of disks, which for the same Debian Mini will cost 3-5p per month. Shelf life is limited only by the amount of money on the balance sheet.

I will not be distracted by prices, they are on the site [1] .

Instead, I’ll tell you about the observations about what numbers are obtained as a result.

First, no matter how surprising, processor time - the least demanded resource (once again that the rent of atoms, which we recently announced , can be quite a reasonable choice) - it accounts for only 5% of the sum of expenses (average consumption was about 6 hours of computer time per week per virtual machine - do not underestimate the power of modern Xeon'ov).
Secondly, the most expensive (and predictable in terms of consumption over time) is RAM - it accounts for about 40% of all costs.
The third in the list is disk storage, which is also quite predictable.

We count all the resources - and the biggest questions are discounting; As far as I know, we are the first to record disk operations for direct disk access (not webDAV or another API).

Disk operation - any request of the OS to read or write information to the disk. A disk operation can be both small and very large (write 4k or write 1Mb - there is clearly a difference), so we divided the accounting of disk operations into two components, reducing the price of each of them - the number of operations is separately considered, the volume of recorded / read data. Of course, the OS cache makes it possible to significantly reduce the amount of disk operations, so far not any access to the web page will cause at least some disk operations. Well, the prices (we ourselves are not yet sure that it will pay off at such prices) are 3 rubles per million disk requests, 10 kopecks per gigabyte of recording.

Based on the general concept, we made the Internet the same way with payment for traffic. Yes, these words cause traumatic memories of the ruble per megabyte and other horrors of the dawn of the Internet. What is paid traffic in the era of the developed Internet and general anlim. But ... If incoming traffic will cost 20 kopecks per gigabyte (outgoing 1p.), Then this is a completely different conversation? After all, if you buy a strip, it implicitly lays in the fact that you consume it at a certain average level. If you consume it under the ceiling, then by purchasing a lane you will already see the real cost of a fully clogged channel. If you do not choose it and 30%, then why are you paying for it? And if you have high outgoing traffic, then why do you implicitly pay for incoming? The second point: what channel do you need? 10 MB? And if it is not enough, if peaks at certain moments are more than 100? We offer gigabit - and the opportunity to pay only for the small part of it that you consume.

And once again about memory. The main feature about which I wrote many months ago is the Memory on Demand technology - allocation of memory on demand. In fact, there is no magic there, we just make sure that the amount of free memory (the cache is not taken into account) remains more or less constant, throwing up memory when there is little free memory and selecting memory when it becomes a lot. Thus, each program gets the required memory capacity - and no more. Of course, this technology is needed only for applications with variable memory consumption (for example, spawning Apaches). In the case of Java / tomcat, the memory should be backed up immediately - and with the question “but what for” it is better to contact Sun Oracle.

This post is a launch announcement. In the next posts I will talk about how this works and why this is the case (short teaser: why 2.6.34-xen, not pv_ops, how the processor time and memory are considered, why we don't have debian lenny, how the system is installed, why Zen 3.4, not 4.01, why XCP, not xend).

Specifically, to lure you into our networks, we set the minimum amount to start working with the cloud at 10p (however, the payment system limits the minimum for replenishing the balance at 100p, it's not us, honestly). Ten rubles will be enough for about three days of messing around with the system. Plus, in our payment model there is no situation when the client owes us. After the end of the funds, we store the disks from the machines for 7 days, if the client decides to continue - he pays for the storage of the disks. If it leaves, there is no court and there is no, we remove the disks and write off their storage as losses. In other words, the cloud test does not commit to anything.