When we see more and more carbon calculators appear to evaluate the impact of "cloud" (even the impact of e-mail or web page display), we want to go back to the source of all these systems by asking ourselves a seemingly simple question: how much does a computer consume, especially a physical server energy consumption calculator?
Intel Xeon processor, 4 cores, 3.5Ghz
32 to 64GB of memory
2 ~ 3 pieces of SDD disks with 256 ~ 512GB.
No graphics card
This server represents most of Dedibox's Pro series, OVH's Advance series, and the current Rise series.
So how much does this server consume? Idle or full? Suppose it is idle, how much does it consume?
A year ago, we began to investigate some system administrators working in data centers and got almost all the answers in the range of 5W to 200W.
It's a bit like buying a gasoline car, and the manufacturer tells you that it consumes 0.1litersr every 100 kilometers. You know in your heart that every 100 kilometers must be between 5 and 15 liters, but you can't be sure.
Similarly, if we don't know what the physical server consumes, we give carbon calculators the ability to announce anything they want. We started with a server, and we didn't know its consumption. Then we added more or less complicated application architecture to it, and finally, we got abstract figures that could not be evaluated.
So, how much does our server consume?
This is not simple.
"measure" electricity consumption
RAPL / Scaphandre
Today, the most commonly used source for measuring server energy consumption calculator is the Intel API called "RAPL" (Running Average Power Limit). This API allows access to the power consumption of the processor and RAM from the operating system. There is little record of how it works; Depending on the generation of the processor, it may be a measurement (with sensors) or an evaluation based on the state of the processor. Sometimes it includes RAM, sometimes it doesn't. Sometimes we even find the global measurement (PSYS) of the whole machine ...
It is worth noting that PowerJoular, Power API, and Scaphandre tools are all based on RAPL.
On our example server above, when the idle RAPL gives us between 3 and 10W, that is less than one LED bulb. Under load, the server energy consumption calculator can reach about 50 or 60W.
However, this only includes CPU and RAM, which is not all. There are also hard drives or SSDs, network cards, motherboards, power supplies, and sometimes RAID cards and GPUs on the server.
electricity meter
Ideally, to measure the total server energy consumption calculator of the server, we need to connect it to an electricity meter. Unfortunately, it is impossible to provide specialized hosting services; As customers, we don't have the freedom to plug in equipment in the data center. Even if we can do it, we need a lot of equipment. An electric meter costs about 20 euros, which is not too bad for an experiment, but we can't install one behind every server. On a larger scale, the cost is even higher, and the cost of rack SmartPDU is much higher.
IPMI
Fortunately, on some servers, we can get the server energy consumption calculator of the power supply through the hardware management interface IPMI. This is not the case with OVH (lack of required cables), but it is the case with Dedibox.
This is good news for us because we have 200 Dedibox servers on hand. We can use them to compare the electricity consumption provided by RAPL/Scaphandre with the overall electricity consumption.
Comparative study: RAPL and Global
In this research, we developed Energizta, a script that can take snapshots of server status every 10 seconds (server energy consumption calculator from RAPL and IPMI, as well as temperature, CPU usage, disk, network, RAM, etc.). Then the average value of these data is calculated every hour and summarized into a database.
This script was executed on more than 200 servers managed by Sysnove for 6 months as a contribution to Boavizta.
In the chart below, each point represents the average value of a one-hour measurement every 10 seconds. We are interested in general trends and orders of magnitude.
Please note that we have chosen to round all measurements to the nearest watt unit. This is enough for our work, and we certainly don't want to give people the illusion of higher accuracy.