Stress Testing Tools? Any Suggestions?

[njcwotx]

I just got me a new toy at work. A 48 core server with 64G RAM and a 16 SAS drive array. Were going to be putting a Database on this thing. I'm used to dealing with dual quads, so this is a new adventure, but before I hand it to my dba I want to get some initial benchmarks.

I am familiar with IOzone for testing disk activity, but iozone may have problems with the amount of ram this thing has. But I wanted to see if there were anyting opensourced tools I could run against the box to put it through a massive load. Something simple that can be loaded and run fairly quickly. Either it can just be a load monster type tool that will stress the box while I use the normal top and iostat and all other normal tools, or it can be a tool that runs and gives me data directly.
_________________
Drinking from the fountain of knowldege.
Sometimes sipping.
Sometimes gulping.
Always thirsting.

[EzInKy]

Would the Phoronix Test Suite do? It's in portage ~arch.
_________________
Time is what keeps everything from happening all at once.

[piwacet]

Not a benchmark tool, but you can run the torture test of mprime (Prime95). I think it will spawn 48 copies and test all your cores. That's what I did with my new rig - it definitely got the CPU temperatures hotter than anything else I've done. Even if I start 50 copies of emerge on my quad core, CPU temperatures are still almost 10 deg C below what mprime torture test caused.

[d2_racing]

[d2_racing]

Can I have a ssh access on that box so that I could run a big fat :

[Moriah]

You had best do a pre-load of all the files first; otherwise you will be choked by the network throughput, or at best the server on the other end with the ebuilds on it you will need to download.

Is that done with 16 ram modules of 4 GB each, and 12 quad cores, all encamped around the ram like the 12 tribes of Israel encamped around the tabernacle in the wilderness?
_________________
The MyWord KJV Bible tool is at http://www.elilabs.com/~myword

Foghorn Leghorn is a Warner Bros. cartoon character.

[pappy_mcfae]

You want a stress test? I've got just the test for you.

Great Internet Mersenne Prime Search. You can install it via portage (emerge -av gimps). Once installed, type mprime. Set the CPU settings to use all the memory you have. Then hit torture test, and wait to see what comes falling out. If your machine can last for 24 hours without failure, you can pretty much guess it will handle anything else you throw at it.

Be advised, this will slow your system to a crawl. I've seen it bring all of my computers to their knees. But they passed the test. You can also help find the next prime number.

Blessed be!
Pappy
_________________
SITE LIST:
Main: http://www.kernel-seeds.org
Mirror: http://kernel-seeds.bloodnoc.org/
Mirror 2: http://kernel-seeds.audiodef.com/
Mirror 3: http://www.elilabs.com/~pappy/

[njcwotx]

So far I have tried the Phoronics Live CD and another bootable tool called Stress Linux.

I do have to laugh at the BogoMIPS aggregate value...230,000. My Core2 workstation runs at an aggregate value of 12,000

Phoronics worked well, but it has a lot of tests and seems to only test performance of specific applications and not the whole box. Most of these tend to test only a thread or two for performance of the specific application for that test and doesn't appear to use all of the box. Havent tested them all though. at no time has any test killed the box there.

Stress Linux basically has a set of tools as well but it has a cmd line tool called 'stress' that you can specify how many threads to run sqrt(), threads for sync(), and threads for malloc()/free() and how many bytes those memory calls request and a few others for hard disks and such.

I ramped up the settings and it seemed like the stress command would hammer the box as much as I wanted. I got the load value in top up to 183 at one point, but it crashed becuase I had it set for 200 threads each requesting 4G of ram so the box froze during that test. I think today Ill play with it a bit more and just try to increase processing threads without over requesting RAM because as long as I did that it didnt seem to mind the load. I didnt have cpu temp monitors running, but I did watch the load % on the 16kva UPS. When I ramped it up it pulled an extra 7% power to run the test

I saw the prime thing but havent run it yet, Ill give that a whirl too.
_________________
Drinking from the fountain of knowldege.
Sometimes sipping.
Sometimes gulping.
Always thirsting.

[njcwotx]

Oh yea, almost forgot to add, I think I do owe it to myself to install a clean copy of gentoo and try to see how long it will compile open office. My Core i7 965 took 30 minutes with 2 RAID 0 SSD drives.

It will be interesting to set my options to compile with 'MAKEOPTS= -j49' and compile on those 16 SAS drives.
_________________
Drinking from the fountain of knowldege.
Sometimes sipping.
Sometimes gulping.
Always thirsting.

[njcwotx]

[Moriah]

[njcwotx]

[Moriah]

So what's the sticker price on a box like that? About $40K I would guess.
_________________
The MyWord KJV Bible tool is at http://www.elilabs.com/~myword

Foghorn Leghorn is a Warner Bros. cartoon character.

[pappy_mcfae]

[njcwotx]

Well I loaded up G2 on the box and learned a few things. Its an opteron job so im using amd64.

I actually build 2 boxes side by side a DL380 and the DL785. The 380 is either 2 Xeons w/hyperthreading or its 2 dual Xeons @ 3.06 (im guessing its the hyperthread job as its a few years old now)

I built the kernels pretty standard, only adding the cciss drivers and checking the processor type and num CPU's, building the kernel from scratch was 6m36s on the DL380 and 1m04s on the DL785.

Emerging the exact same set of packages afterwards while still in the chrooted environment was 14:27 on the DL785 and 18:19 on the DL385.

After the first reboot to the build kernel and finalizing the little things, I emerged Gnome. On the DL380 it was 498m58s, and on the DL 785 it took 382m43s. On a side note, I watch both machines with top and vmstat almost the whole time and the Xeon box was blazing in the upper 20% of total CPU while the 48core job was barely hitting 1-2% with the very occaisional jump to 12%. So even though times were not significantly different, the servers were completely different in response.

________________________
Got me to thinking about multicores a bit. in this case the CPU speeds are 3.06 Xeon w HT(x2) vs 2.4 Opteron(x48) mhz. So an single thread will run a bit faster on the Xeon. One might expect to see blazing speeds on builds but I guess it does stand to reason that your -jxx setting helps to a point. because that only affects the compiling portions. I did prefetch the files, but theres a lot of other things besides compiling going on too, and some packages might only need one thread or two to build. To my knowlege an emerge only works one package at a time, so if a build is one big fat binary and uses one thread, no other cpu's are getting any requests to work.

Is there a way to make emerge work on more than one package at once? I imagine if I can get the 400 packages to compile at 10 per job it may go much faster on the bigger box.
_______________________

after adding all the extra cpu's and hdd's and necessary guts to put it all in it was closer to 55 Id say.

_______________________
--couple other notes, I mentioned office took 28 mintues to build, it might have been 2hrs28mins! but I can't remember. Now im wondering just how long it took. I had a single core take 12 hours once, I just remember being impressed with i7 and raid 0 ssd's

--didnt realize it but the stress linux distro uses mostly if not all public tools, the package is called 'stress' and is in portage along with a few others. Here the link to the list of tools. http://www.stresslinux.org/sl/wiki/Software

--more testing tommorrow, before too long I have to wrap this up and let someone else have all the fun.
_________________
Drinking from the fountain of knowldege.
Sometimes sipping.
Sometimes gulping.
Always thirsting.

[Dancingfire]

you can use portage-2.2, it supports --jobs, and --load-average
give it a try


as a matter of "stress testing" you can also try compiling kernel with -j (without any number).
smth like 20 rounds of make -j && make clean will be nice)