Slashdot Mirror
Quick, Standard Measurement for CPU Power?
captnitro asks: "A particular research project I'm developing right now needs to compare 'potential' (idle/none) and 'load' for various hardware capabilities, and quickly -- maybe up to a several times every minute. For disk space, for RAM, it's relatively easy -- find what's used and what's not and report the ratio. For CPU, I have plenty of time to test 'potential' when the app starts. But for testing CPU load, I need a standard 'ruler' that will be able to compare across varying platforms and processors (e.g., x86, PowerPC, embedded, single and multi-proc) -- so for example, idle percentage won't work. At the same time, I don't have the ability to time 'openssl speed' every 25 seconds without bringing the system to a halt. I'm willing to sacrifice precision of the measurement for generalization of the unit -- that is, the operations that this test is for would be primarily mathematical and not say, text sorts -- but I'd prefer a generic, quick test of the current processor load rather than an average of 25 different tests. Regardless of hardware, the OS distribution is mostly *nix-based -- NetBSD, Linux, and even Mac OS X. Wild ideas are perfectly acceptable -- any thoughts?"
everyone understands bogomips, right?
I am not even sure there is a way of measuring what you want. There are so many variables (Disk I/O, memory bandwidth, etc.) that you can't get a reasonable measurement of how much CPU is left over to do useful work for your process.
The other thing is that many things can take down a CPU as well - a huge burst of network traffic can take an "idle" CPU and peg it at 100% in kernal usage.
Your best bet would probably be to hook into the various schedulers and get deep knowledge of what is really going on - pain in the butt, but doable
I have mod points and I am not afraid to use them
Watts. http://www.convert-me.com/en/convert/units/power/p ower.kgfmpersec.en.html
Mhz? :)
Or!
Watts of thermal power on the die surface!
idle percentage won't work
Why not? Just normalize it against total CPU capacity (MIPS, GFLOPS, or whatever). I'm probably not explaining it right, but I really don't see this as a difficult problem to solve.
Hmm... maybe MIPS, GFLOPS, etc is the "ruler" you're looking for. If a processor can bechmark at 10GFLOPS (running at 100%), then a machine 90%idle is performing a maximum of 1GFLOP worth of work. It's sloppy as hell, but it might be good enough.
Run a dummy task at an extremely low priority. A bechmark of that task at the lowest possible priority will give you a good indication of how much CPU power was 'left over'. This will only work if you can make the dummy task a lower priority than everything else in the system.
just use a voltmeter you idoit
It sounds like you are looking for a way to normalize and compare system (and application?) performance across different hardware platforms.
Say hello to SPEC. This is exactly what the organization was formed for. Take a look at their CPU benchmarks. I know you're looking for more of a snapshot, and less of a benchmark - but I would think SPEC is a good place to start.
Perhaps I am misunderstanding the question, but could you measure current flow to the CPU to measure its operating level? You know the idle current flow, and you know the expected peak draw, so by measuring current current flow through the processor you should be able to tell how loaded it is.
Degaussing scares the bad magnetism out of the monitor and fills it with good karma.
You seem to hint that openssl speed is the type of metric you want... So why not run openssl speed once and use one of the tests as your power rating. For example say you choose blowfish cbc 8192 as the power rating. Total power for group of servers would be the sum of all the power ratings. Available processing power would server cpu idle % * power rating. A few simple scripts and you could have a rough approximation of available processing power available across all platforms. You could also design a benchmark based on the type of processing you will be doing, to make it more relevant.
One Mississippi.
Two Mississippi.
Three Mississippi.
Whatever you do, give energy consumption some weight in the comparison. In this day and age, low energy requirements is a virtue. In other words, for two CPUs that are otherwise equivalent, the one which consumes lower energy ought to win.
must... stay... awake...
Quick
Fast
Freakin Fast
Holy Shit
Dude you r0x0r
OMFG
Oh, just so you know, you should always buy at the Holy Shit level. If you buy OMFG you pay too much and you will soon regret buying anything less than Freakin Fast. But Holy Shit and you will be happy for a couple to three years without paying too much.
The paper you're looking for is Wong, Brian, "Comparing MVS and UNIX Workload Characteristics" Int. CMG Conference, 1998, in which (if memory serves) he looks at the comparability of MVS and Unix programs, and derives a series of comparisons which very approximately correlate with TP performance in a benchmark like TPC.
A paper that's available to CMG members is his Developing a General-Purpose OLTP Sizing Tool, which builds on the subject.
If you have any good, representative test that load a whole system, then in my opinion, you have a good predictor of average performance. (Tautology, eh?) If you have such a test, though, people will learn how to get the best posible numbers from it. Average together all the various X MHz Pentium III TPC-C results, don't trust any single one (:-))
--dave
davecb@spamcop.net
Setup a low end pc with a NIC and OS of your choice. Setup a share on the server to be tested. On your low end pc, write a script or "code" to move a small file to the server and then move it back. Log the rates, times and delta comparisons on the low end PC. Be sure to give the file a unique name each move. Something like the time stamp should work. You could, with the proper coding, derive a good bit of data without adversely affecting the test subject.
catch a wave and ride till ya hit the beach
Consider:_ per_second
http://en.wikipedia.org/wiki/Flops
http://en.wikipedia.org/wiki/Million_instructions
When you're afraid to download music illegally in your own home, then the terrorists have won!
1) do some research on figuring out the relationship between Amps (current draw) and units of computation per time (like flops maybe). I.e., is it a linear relationship? Come up with a relationship.
2) Using #1, calculate the current draw of CPU (perhaps use a separate dedicated power supply for the CPU and/or mobo) and therefore the current CPU utilization.
You might need to address temperature in there too. Or *only* use temperature rather than current (for instance temperature difference between the outside of the case and the CPU).
Just an idea. Some CPUs are more efficient than others so you have to scale it, but doesn't it make sense that CPU power is proportional to real power?
http://the1.no-ip.com/~the1/johnbench.txt
You can download the john the ripper source code here:
http://www.openwall.com//john/
and then build it and run your own tests.
------ Take away the right to say fuck and you take away the right to say fuck the government.
Watts?
Not enough information was given in the question, but it is possible that the application in mind is that of measuring the performance of webservers in real time. So the experimenter doesn't (or does?) want the tests to last the long time a good SPEC run would take. I was somewhat confused by the article text.
Then again, it's possible that the application could be used to measure the speed of web clients. In that case, a flash or javascript loop with whirling icons could do the trick, but it would take some work. Since it's possible that this may be used (either now or in the future) by marketers who want to measure the comparative speed of a computer (to use to categorize someone into a marketable demographic so they can "target" the right ads), maybe answering this question fully wouldn't be such a good idea.
"...needs to compare 'potential' (idle/none) and 'load' for ... CPU ... test 'potential' when the app starts."
... 'ruler' ... compare ... x86, PowerPC, embedded, single and multi-proc ..."
... mostly *nix-based"
what is the app? is it your app, only one app, could be any one of hundreds of apps?
"need
"idle percentage won't work"
because you want to know how much work you can do
"sacrifice precision of the measurement for generalization of the unit"
"generic, quick test of the current processor load
"thoughts?"
standard guage, just measure the right stuff
if all apps belong to you, just calibrate according to results rate...
full = no aps running therefore no rate
empty = all apps running therefore full rate
guage = empty rate - current rate
That's $20 + % gross rev, if it works...
CPUs aren't comparable, so you need to get a % and normalize it.
The easiest way to get a % is from something like top - but you can definitely do better, especially by successfully taking into account priority. You definitely want to find a way to _access_ information the scheduler already has, not make little minibenchmarks.
But you still have the problem of 100% of a PII vs 100% of a G5. This is a principally insoluble problem that varies based on your application's use of int, flop, registers, L1, L2, memory bandwidth and offloading chips (such as better NICs)
An ok solution would be to read the CPU class and do a lookup in a table that you keep upgrading as you get better info. My shorthand table is at the bottom of this post.
A better solution would be to do a relevant CPU-intensive benchmark on each machine and store that result. If what you care about is OpenSSL speed, that's the best benchmark FOR YOU.
The best solution is to run a series of full-setup benchmarks. Run it at high, middle and low memory usage, since you're measuring that separately, etc... then do all combination with high/low disk bandwidth. Run it at high/med/low CPU based on whatever % measure you settled on. You'll get not just a number but a "map" of performance based on other factors. If this matrix is very detailed it might automatically perform hundreds of combinations and take a long time once, but you'd get very accurate results.
I'm interested to see your results, but until you take them, this is my ruthlessly shorthand table which disregards huge amounts of things and wildly approximates all sorts of things and some stuff I've made up:
P4 CPU speed x1
P3 CPU speed x1.2
P2 CPU speed x.8
Athlon XP xxxx+ rated speed.
G3 CPU speed x1.8
G4 CPU speed x1.5
G5 CPU speed x2
Dual CPU x1.7 (probably dual core also)
Quad CPU x3 (probably dual/dual cpu/core also)
(please note: PIIIs definitely were often clock-for clock faster than early PIVs, at least. But they couldn't scale up. Same with G3 vs G4. Note that chess is highly integer and nothing is better than an Intel, apparently.)
Looking for freelance Actionscript (Flash/Flex) or ColdFusion work and/or freelance developers. Email me, put Slashdot