Energy consumption vs % utilization?
Sorry, this is somewhat OT. I'm looking for information on energy consumption vs percent utilization. In other words if your datacenter consumes 720 MWh per month, yet on average your servers are 98% underutilized, you are wasting a lot of energy (a hot topic these days). Does anyone here have any real data on this? Grisha
On Tue, Oct 26, 2004 at 01:52:51PM -0400, Gregory (Grisha) Trubetskoy wrote:
Sorry, this is somewhat OT.
Also Sorry, but I think the question itself is completely flawed.
I'm looking for information on energy consumption vs percent utilization. In other words if your datacenter consumes 720 MWh per month, yet on average your servers are 98% underutilized, you are wasting a lot of energy (a hot topic these days). Does anyone here have any real data on this?
What does 98% underutilized mean? What is the utilization of a device with fully built out RAM that is used to 100%, when the CPU is used 2% only? What is the utilization of a system, that uses two percent of the memory and two percent of the available CPU time, when the policy of the top secret organization owning this system requires, that the application is running on a seperated machine? Sure many machines might be (computing power wise) able to handle Firewalling, Routing, Webserving, Database Serving, Mailserving and storing accounting data, but still there might be very good reasons to seperate these on different machines. If you take points like policy requirement (see above: an application might by policy utilize a machine to 100%), different types of resources, failover etc. into account, you might end up with different numbers then just looking at the CPU (and I have the feeling that is what you did or were intending to do). Actually I think nobody does calculate "real" utilization, as there are a lot of soft factors to be taken into account. Nils
Actually I think nobody does calculate "real" utilization, as there are a lot of soft factors to be taken into account.
Electrical usage for a datacenter is pretty consistent throughout a month, even as measured by a sum of days. The utilization of the systems inside of it are almost anything but consistent... even during boot up it would be nearly impossible to determine the instantaneous necessary power draw. Separately, deploying applications to clusters of machines where the cluster is dynamically resized [more machines are turned on/off] depending on load is a non-trivial function and outside the operational experience/need of most customers. But even assuming you could do that, the best approximation I could imagine for an Internet data center would be something akin to its network traffic graph [assumption being that network load amongst a stable set of customers is proportionate to the processing power required to produce it... even if an individual customer uses much more CPU power to do that at a specific time quanta]. Basically, if you use 1Mb/s at noon on Monday, and 1.2Mb/s at noon on Tuesday with the same customer set, you can probably estimate that your system's load is 20% higher than it was on Monday. Assuming you aren't operating at either the very low extreme or very high extreme. At least that would be my thought. If all applications were designed to virtualized ala mainframe style, this clustering concept might work to dynamically redeploy resources... However the mainframes themselves are inherently not-smooth-stepped in terms of their power/cpu curves, so its probably a dead issue in that regard. Deepak Jain AiNET
On Tue, 2004-10-26 at 19:52, Gregory (Grisha) Trubetskoy wrote:
In other words if your datacenter consumes 720 MWh per month, yet on average your servers are 98% underutilized, you are wasting a lot of energy (a hot topic these days).
Which means you have to make sure the revenue generated by those 98% underutilized servers covers your powerbill and other expenses, preferrably leaving some headroom for a healthy profit margin. As long as that's the case there's no real waste of energy, the services people run on their servers are supposed to be worth the energy and other costs, whether they physically fully utilize their power or not. Cheers, -- --- Erik Haagsman Network Architect We Dare BV tel: +31(0)10 7507008 fax:+31(0)10 7507005 http://www.we-dare.nl
Erik Haagsman wrote:
Which means you have to make sure the revenue generated by those 98% underutilized servers covers your powerbill and other expenses, preferrably leaving some headroom for a healthy profit margin. As long as that's the case there's no real waste of energy, the services people run on their servers are supposed to be worth the energy and other costs, whether they physically fully utilize their power or not.
Yet there are a lot of clusters which are designed for peak load, which will waste energy during non-peak hours. Developing an in-house system for shutting down power to excess servers in a cluster might increase the healthy profit margin. -Jack
Hello, I've done quite a bit of studyin power usage and such in datacenters over the last year or so.
I'm looking for information on energy consumption vs percent utilization. In other words if your datacenter consumes 720 MWh per month, yet on average your servers are 98% underutilized, you are wasting a lot of energy (a hot topic these days). Does anyone here have any real data on this?
I've never done a study on power used vs. CPU utilization, but my guess is that the heat generated from a PC remains fairly constant -- in the grand scheme of things -- no matter what your utilization is. I say this, because, with a CPU being idle of 100% utilized, they still are grossly inefficient, on the order of less than 10% in all cases (ie, 1 watt in returns at least .9 watts of heat, no matter loading of the CPU). -- Alex Rubenstein, AR97, K2AHR, alex@nac.net, latency, Al Reuben -- -- Net Access Corporation, 800-NET-ME-36, http://www.nac.net --
In message <Pine.WNT.4.61.0410261429110.3340@vanadium.hq.nac.net>, Alex Rubenst ein writes:
Hello,
I've done quite a bit of studyin power usage and such in datacenters over the last year or so.
I'm looking for information on energy consumption vs percent utilization. In
other words if your datacenter consumes 720 MWh per month, yet on average your servers are 98% underutilized, you are wasting a lot of energy (a hot topic these days). Does anyone here have any real data on this?
I've never done a study on power used vs. CPU utilization, but my guess is that the heat generated from a PC remains fairly constant -- in the grand scheme of things -- no matter what your utilization is.
I doubt that very much, or we wouldn't have variable speed fans. I've monitored CPU temperature when doing compilations; it goes up significantly. That suggests that the CPU is drawing more power at such times. Of course, there's another implication -- if the CPU isn't using the power, the draw from the power line is less, which means that much less electricity is being used. --Steve Bellovin, http://www.research.att.com/~smb
I doubt that very much, or we wouldn't have variable speed fans. I've monitored CPU temperature when doing compilations; it goes up significantly. That suggests that the CPU is drawing more power at such times.
I don't doubt what you are saying. However, I did say, "in the grand scheme of things", meaning that the heat given off by the CPU, and change thereof, relative to the constant heat given off by the rotation of hard drives, the heat given off by the power supplies, etc., is still small.
Of course, there's another implication -- if the CPU isn't using the power, the draw from the power line is less, which means that much less electricity is being used.
An important point, but I still bet relatively small. It's going to be a busy weekend at the Rubenstein Lab (aka, my garage) this weekend; I'll post results to my findings. -- Alex Rubenstein, AR97, K2AHR, alex@nac.net, latency, Al Reuben -- -- Net Access Corporation, 800-NET-ME-36, http://www.nac.net --
Steven M. Bellovin wrote:
In message <Pine.WNT.4.61.0410261429110.3340@vanadium.hq.nac.net>, Alex Rubenst ein writes:
Hello,
I've done quite a bit of studyin power usage and such in datacenters over the last year or so.
I'm looking for information on energy consumption vs percent utilization. In
other words if your datacenter consumes 720 MWh per month, yet on average your servers are 98% underutilized, you are wasting a lot of energy (a hot topic these days). Does anyone here have any real data on this?
I've never done a study on power used vs. CPU utilization, but my guess is that the heat generated from a PC remains fairly constant -- in the grand scheme of things -- no matter what your utilization is.
I doubt that very much, or we wouldn't have variable speed fans. I've monitored CPU temperature when doing compilations; it goes up significantly. That suggests that the CPU is drawing more power at such times.
From running a Colo in a place with ridiculus high electricity engery costs (Zurich/Switzerland) I can tell you that the energy consuption of routers/telco (70%) and servers (30%) changes changes significantly throughout the day. It pretty much follows the traffic graph. There is a solid base load just because the stuff is powered up and from there it goes up as much as 20-30% depending on the routing/computing load of the boxes. To simplify things you can say that per packet you have that many "mWh" (milli-Watt-hours) per packet switched/routed or http requests answered over the base load. I haven't tried to calulate how much energy routing a packet on a Cisco 12k or Juniper M40 cost though. Would be very interesting if someone (student) could do that calculation. -- Andre
From: "Andre Oppermann"
From running a Colo in a place with ridiculus high electricity engery costs (Zurich/Switzerland) I can tell you that the energy consuption of routers/telco (70%) and servers (30%) changes changes significantly throughout the day. It pretty much follows the traffic graph. There is a solid base load just because the stuff is powered up and from there it goes up as much as 20-30% depending on the routing/computing load of the boxes. To simplify things you can say that per packet you have that many "mWh" (milli-Watt-hours) per packet switched/routed or http requests answered over the base load. I haven't tried to calulate how much energy routing a packet on a Cisco 12k or Juniper M40 cost though. Would be very interesting if someone (student) could do that calculation.
the same variation between night and day here - but from our point of view the consumption of the air-pack's are making the differences during the day ... traffic-graph and outside temperature-graphs show more or less the same up and down. Would we interesting to have separate values for the power consumption between server-equipment and air-co ... Greetings Nik
Nik Hug wrote:
From: "Andre Oppermann"
From running a Colo in a place with ridiculus high electricity engery costs (Zurich/Switzerland) I can tell you that the energy consuption of routers/telco (70%) and servers (30%) changes changes significantly throughout the day. It pretty much follows the traffic graph. There is a solid base load just because the stuff is powered up and from there it goes up as much as 20-30% depending on the routing/computing load of the boxes. To simplify things you can say that per packet you have that many "mWh" (milli-Watt-hours) per packet switched/routed or http requests answered over the base load. I haven't tried to calulate how much energy routing a packet on a Cisco 12k or Juniper M40 cost though. Would be very interesting if someone (student) could do that calculation.
the same variation between night and day here - but from our point of view the consumption of the air-pack's are making the differences during the day ... traffic-graph and outside temperature-graphs show more or less the same up and down. Would we interesting to have separate values for the power consumption between server-equipment and air-co ...
In this case the air-co is not included. That is measured on a separate circuit for which I don't have any figures ready. Also note that especially high-end routers draw power load dependent. With SONET/SDH stuff I haven't seen it. The reason is circuit switching. They switch continuously the same amount of data. -- Andre
Alex Rubenstein wrote:
I'm looking for information on energy consumption vs percent utilization. In other words if your datacenter consumes 720 MWh per month, yet on average your servers are 98% underutilized, you are wasting a lot of energy (a hot topic these days). Does anyone here have any real data on this?
I've never done a study on power used vs. CPU utilization, but my guess is that the heat generated from a PC remains fairly constant -- in the grand scheme of things -- no matter what your utilization is.
You should be able to pick up simple current / wattage meter from local hardware store for $20 or so. That will tell you that on a modern dual-CPU machine the power consumption at idle CPU is about 60% of peak. The rest is consumed by drives, fans, RAM, etc. As wattage the difference is 100-120W (50-60W per cpu) All modern operating systems do moderate job of saving CPU wattage when they are idle (BSD's, Linux, MACOS X, WinXP, etc.) Pete
Speaking on Deep Background, the Press Secretary whispered:
You should be able to pick up simple current / wattage meter from local hardware store for $20 or so. That will tell you that on a modern dual-CPU machine the power consumption at idle CPU is about 60% of peak. The rest is consumed by drives, fans, RAM, etc. As wattage the difference is 100-120W (50-60W per cpu)
Bogus data alert.... A ammeter will tell you amps. But in the world of switcher power supplies, that does not beget watts. [Why? is an exercise for the student. Start with "power factor" and "VARS" and worry about asymetric loads...] If you want to talk watts, as you must to worry about HVAC, or really watt-hours... acquire a power company type meter - in glass with a whirlygig. Put it in a meter box with plugs. [See your local eletrical wholesaler..] (The rotating disk watthour meter is amazingly accurate under almost any kind of load waveform. Only time-of-day metering has spurred the utilities to replace them.) Plug the machines into it; it into the wall. Note the numbers and the time, and come back in 24Hours. -- A host is a host from coast to coast.................wb8foz@nrk.com & no one will talk to a host that's close........[v].(301) 56-LINUX Unless the host (that isn't close).........................pob 1433 is busy, hung or dead....................................20915-1433
On Tue, 26 Oct 2004, David Lesher wrote:
Speaking on Deep Background, the Press Secretary whispered:
You should be able to pick up simple current / wattage meter from local hardware store for $20 or so. That will tell you that on a modern dual-CPU machine the power consumption at idle CPU is about 60% of peak. The rest is consumed by drives, fans, RAM, etc. As wattage the difference is 100-120W (50-60W per cpu)
Bogus data alert....
A ammeter will tell you amps. But in the world of switcher power supplies, that does not beget watts. [Why? is an exercise for the student. Start with "power factor" and "VARS" and worry about asymetric loads...]
If you want to talk watts, as you must to worry about HVAC, or really watt-hours... acquire a power company type meter - in glass with a whirlygig. Put it in a meter box with plugs. [See your local eletrical wholesaler..]
(The rotating disk watthour meter is amazingly accurate under almost any kind of load waveform. Only time-of-day metering has spurred the utilities to replace them.)
Plug the machines into it; it into the wall. Note the numbers and the time, and come back in 24Hours.
Instead of doing all this, just buy a Kill-A-Watt meter for about $30, and get an instant reading of Watts, Amps, VAs, power factor, and KWH. K
Speaking on Deep Background, the Press Secretary whispered:
[KWH meter]
Instead of doing all this, just buy a Kill-A-Watt meter for about $30, and get an instant reading of Watts, Amps, VAs, power factor, and KWH.
Interesting. I've not seen anything near that cheap. The spec sheet is rather lacking but if it does the job.. Note that it's 120V only and 15A; lots of racks exceed that. -- A host is a host from coast to coast.................wb8foz@nrk.com & no one will talk to a host that's close........[v].(301) 56-LINUX Unless the host (that isn't close).........................pob 1433 is busy, hung or dead....................................20915-1433
On Tue, 26 Oct 2004 13:52:51 EDT, "Gregory (Grisha) Trubetskoy" said:
average your servers are 98% underutilized, you are wasting a lot of
Remember in your analysis to include premature hardware failure due to too many power cycles... A server can *easily* "on average" be running at only 20-30% of capacity, simply because requests arrive at essentially random times - so you have to deal with the case where "average" over a minute is 20% of capacity for 600 hits (10/sec), but some individual seconds only have 1 hit, and others have 50 (at which point you're running with the meter spiked). Time-of-day issues also get involved - you may need to have enough iron to handle the peak load at 2PM, but be sitting mostly idle at 2AM. Unfortunately, I've seen very few rack-mount boxes that support partial power-down to save energy - if it's got 2 Xeon processors and 2G of memory, both CPUs and all the memory cards are hot all the time... There's also latency issues - if some CPUs on a node or some nodes in a cluster are powered down, there is a timing lag between when you start firing them up and when they're ready to go - so you need to walk the very fine line between "too short a spike powers stuff up needlessly" (very bad for the hardware), and "too much dampening means you get bottlenecked while waiting for spin-up". (Been there, done that - there's a 1200-node cluster across the hall, and there's no really good/easy way to ramp up all 1200 for big jobs and power down 800 nodes if there's only 400-nodes worth of work handy. So we end up leaving it all fired up and let the node's "idle loop" be "good enough").. If it was as easy as all that, we'd all be doing it already.. :)
Sorry, this is somewhat OT.
I'm looking for information on energy consumption vs percent utilization. In other words if your datacenter consumes 720 MWh per month, yet on average your servers are 98% underutilized, you are wasting a lot of energy (a hot topic these days). Does anyone here have any real data on this?
Grisha
It is only waste is the P & L statement is showing no profit.
Thats an insane statement. Are you saying, "You are only wasting money on things if you aren't profitable" ? /action shakes head. On Tue, 26 Oct 2004, james edwards wrote:
Sorry, this is somewhat OT.
I'm looking for information on energy consumption vs percent utilization. In other words if your datacenter consumes 720 MWh per month, yet on average your servers are 98% underutilized, you are wasting a lot of energy (a hot topic these days). Does anyone here have any real data on this?
Grisha
It is only waste is the P & L statement is showing no profit.
-- Alex Rubenstein, AR97, K2AHR, alex@nac.net, latency, Al Reuben -- -- Net Access Corporation, 800-NET-ME-36, http://www.nac.net --
It's more or less the truth though. Only on rare occasions, such as the cluster/fail-over scenario given, can you actually supply less power to certain machines, and power use largely unrelated to their actual utilisation. Keep an eye on your UPS load during peak hours and you'll see the load rising when traffic and server utilisation rises, but compared to the baseline power needed to feed servers these fluctuations are peanuts. You supply a server with enough power to run...how is this waste exactly...? If anyone is wasting anything, it's perhaps hardware manufacturers that don't design efficiently enough, but power that you provide and that's used (and paid for) by your customers is not wasted IMO. Cheers, Erik On Tue, 2004-10-26 at 21:07, Alex Rubenstein wrote:
Thats an insane statement.
Are you saying, "You are only wasting money on things if you aren't profitable" ?
/action shakes head.
On Tue, 26 Oct 2004, james edwards wrote:
Sorry, this is somewhat OT.
I'm looking for information on energy consumption vs percent utilization. In other words if your datacenter consumes 720 MWh per month, yet on average your servers are 98% underutilized, you are wasting a lot of energy (a hot topic these days). Does anyone here have any real data on this?
Grisha
It is only waste is the P & L statement is showing no profit.
-- Alex Rubenstein, AR97, K2AHR, alex@nac.net, latency, Al Reuben -- -- Net Access Corporation, 800-NET-ME-36, http://www.nac.net -- --
Erik Haagsman Network Architect We Dare BV tel: +31(0)10 7507008 fax:+31(0)10 7507005 http://www.we-dare.nl
On Tue, 26 Oct 2004, Erik Haagsman wrote:
It's more or less the truth though.
I think the comment was outside of the scope of the original discussion. It seemed to me that:
It is only waste is the P & L statement is showing no profit.
inferred that any business practice is OK, as long as your are profitable. It is that concept that I felt was insane. -- Alex Rubenstein, AR97, K2AHR, alex@nac.net, latency, Al Reuben -- -- Net Access Corporation, 800-NET-ME-36, http://www.nac.net --
Thats an insane statement.
Are you saying, "You are only wasting money on things if you aren't profitable" ?
/action shakes head.
No, I am not but my statement did sure sound like that was what I was saying. I do think it is apples or oranges comparing CPU % to total power used and coming up with a wasted factor. My colo needs X amps/hour just to run at idle, I don't call this waste. It is the cost of doing business. Power factor causes losses. So you need enough customers to cover this and other expenses. I guess we need a definition of waste here. I would say the the heat produced by pulling all the amp/hrs is waste. It could be possible to harvest this and reuse it elsewhere. So, just because you are profitable does not mean there is no waste but it also depends on how you classify waste. Also, do the methods to avoid this waste justify (pay for over time) their use. James H. Edwards Routing and Security Administrator At the Santa Fe Office: Internet at Cyber Mesa jamesh@cybermesa.com noc@cybermesa.com http://www.cybermesa.com/ContactCM (505) 795-7101
participants (14)
-
Alex Rubenstein
-
Andre Oppermann
-
David Lesher
-
Deepak Jain
-
Erik Haagsman
-
Gregory (Grisha) Trubetskoy
-
Jack Bates
-
james edwards
-
Krzysztof Adamski
-
Nik Hug
-
Nils Ketelsen
-
Petri Helenius
-
Steven M. Bellovin
-
Valdis.Kletnieks@vt.edu