FLOPs on a guest virtual machine differ from those of a host machine??


Message boards : Windows : FLOPs on a guest virtual machine differ from those of a host machine??

Message board moderation

To post messages, you must log in.
AuthorMessage
candido

Send message
Joined: 28 Jul 12
Posts: 21
Credit: 2,513,800
RAC: 0
Message 245 - Posted: 16 Sep 2012, 9:41:00 UTC

Last modified: 16 Sep 2012, 10:11:07 UTC
Hi,after pondering i decided to give it a try and installed Oracle VirtualBOX, then Ubuntu 12.04 on a virtual machine, and then linux BoincManager on this virtual machine.
I followed the instructions here (http://asteroidsathome.net/boinc/forum_thread.php?id=27&nowrap=true#166) for preparing the VBOX CD-ROM.
Then followed the wizzard recomendations of the oracle VBox instalation program. (Did not change any of the recommended / default instaltion values for the creation of a virtual machine , including type of virtual hard drive, size, size of virtual RAM, etc.)
After that I followed these instructions (http://asteroidsathome.net/boinc/forum_thread.php?id=18&nowrap=true#212) to install boincmanager on my linux virtual machine. I got version 7.0.36.

My computer is a 64bit laptop with approx 3 or 4 years, Pentium(R) Dual-Core CPU T4300 @ 2.10GHz [Family 6 Model 23 Stepping 10]. The Host OS is Win 7, 64 bit.
It has been showing consistently on other project I am participating:

Measured floating point speed: 2131.79 million ops/sec
Measured integer speed: 6294.65 million ops/sec

But the virtual machine I installed (Linux 3.2.0-29-generic, BOINC 7.0.36, RAM 495.97 MB, Cache 6144 KB, Swap space 510 MB, Total disk space 7.47 GB, Free 4.64 GB) shows only:

Measured floating point speed: 1452.09 millão ops/seg (-32%)
Measured integer speed: 4872.34 millão ops/seg (-23%)

This is very big drop in the computing capacity. Is this real? Has there been a reall drop in computing capacity in the virtual machine?

Additional info:
1. If i run CPU benchmarks, numbers vary a bit but not significantly.

2. I still have boinc manager running on the host OS (windows 7) but am using only 50% of the processors (meaning 1 of the existing 2).

3. before installing the virtual machine I isntalled ubuntu side-by-side with windows twice, and twice had problems. Boinc did not run appropriatly, etc.
But on these two ocasions, the same laptop showed these numbers:

1st instalation
Linux 2.6.32-41-generic-pae
BOINC 6.10.17
RAM 3859.25 MB
Cache 1024 KB
FLOP : 1354.12 millão ops/seg
Measured integer speed: 4807.79 millão ops/seg

2nd install
Linux 3.2.0-27-generic
BOINC 7.0.27
RAM 3794.28 MB
Cache 1024 KB
FLOP: 2093.4 millão ops/seg
Measured integer speed: 10950.69 millão ops/seg (this time it went up!?)

My question is, by using a virtual machine am I losing crunching power?
Would it be bigger if I used both processors on my host OS (on other projects) and none in the virtual machine (on asteroids)?
thanks

ccandido

Edit: I just got my first 120 credits on this projec, the WU ran for about 6,5 hours:
48302 18913 16 Sep 2012 | 2:51:35 UTC 16 Sep 2012 | 9:51:34 UTC Completo e validado 23,704.59 21,126.57 120.00 Period Search Application v101.00
ID: 245 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Kyong
Project administrator
Project developer
Project tester
Project scientist
Avatar

Send message
Joined: 9 Jun 12
Posts: 584
Credit: 52,667,664
RAC: 0
Message 248 - Posted: 16 Sep 2012, 13:22:55 UTC
Hi, there should be no losing power by using a virtual machine. I have full computing power in virtual machines. I've run many tests but I can't much tell about your configuration. I use mainly Debian and I have OpenSuse and Gentoo on some other computers. I really don't like Ubuntu because of many things so I haven't much experience about computing power in BOINC on it. If you can try to use Debian and some stable version of BOINC. Version 7.0.36 is testing release, not stable, so there may some problems, maybe about benchmarks, because your computing time of the unit (6.5 hours) seems allright on the type of processor.
ID: 248 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
candido

Send message
Joined: 28 Jul 12
Posts: 21
Credit: 2,513,800
RAC: 0
Message 249 - Posted: 16 Sep 2012, 14:03:33 UTC - in response to Message 248.  
Thanks Kyong.
In case I decide to install a virtual machine on other computers I will try to use Debian and a more stable version of Boinc Manager and then check the number of Flops again. In case they don't show a big diference I will replace ubuntu with Debian. Meanwhile i am almost completing my 2nd asteroids WU and so far so good, And, since you mentioned that the 6,5 hours processing time for a WU seems OK, I will leave this machine untouched for a while and see how it goes.
ccandido
ID: 249 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
jujube

Send message
Joined: 5 Sep 12
Posts: 30
Credit: 24,320
RAC: 0
Message 251 - Posted: 16 Sep 2012, 20:48:10 UTC
I have observed the same as ccandido... the benchmarks from BOINC running on a virtual Ubuntu machine are considerably lower than normal BOINC benchmarks. I can't understand why they would be lower on a Ubuntu VM but not a VM built from a different Linux distro but I intend to verify that claim. I don't like Ubuntu either so maybe it's time to investigate and perhaps find a better distro. Or maybe Ubuntu needs to be configured a certain way when installed as a VM. Maybe folks at the Ubuntu forums have some advice on the subject.
ID: 251 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
fractal

Send message
Joined: 12 Jul 12
Posts: 5
Credit: 20,198,079
RAC: 0
Message 254 - Posted: 16 Sep 2012, 22:58:42 UTC - in response to Message 245.  
2. I still have boinc manager running on the host OS (windows 7) but am using only 50% of the processors (meaning 1 of the existing 2).

This is probably why you are getting inaccurate benchmarks.

Stop boinc and anything else that is processor intensive in windows before running benchmarks in the VM if you want accurate benchmark results.
ID: 254 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
jujube

Send message
Joined: 5 Sep 12
Posts: 30
Credit: 24,320
RAC: 0
Message 256 - Posted: 17 Sep 2012, 4:56:15 UTC - in response to Message 254.  
Stop boinc and anything else that is processor intensive in windows before running benchmarks in the VM if you want accurate benchmark results.


That doesn't help, I still get poor benchmarks in the VM. Normal benchmarks for my i7 2600K at stock clock (3.40 GHz) are 3,770 and 22,028. In the Ubuntu VM benchmarks are 3,227 and 13,884. Wow, a huge hit on the IOPs.

I installed 32 bit Ubuntu in the VM along with XFCE desktop because that combination seemed to make the system more stable and responsive. Does 32 bit BOINC benchmark lower than 64 bit BOINC?

I'm burning a 32 bit Fedora 17 Live DVD as I post this. I'll report the benchmarks as soon as I get them. I might have to try a 64 bit Fedora 17 VM for comparison.

ID: 256 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
candido

Send message
Joined: 28 Jul 12
Posts: 21
Credit: 2,513,800
RAC: 0
Message 258 - Posted: 17 Sep 2012, 8:52:08 UTC - in response to Message 256.  
Stop boinc and anything else that is processor intensive in windows before running benchmarks in the VM if you want accurate benchmark results.


That doesn't help, I still get poor benchmarks in the VM. Normal benchmarks for my i7 2600K at stock clock (3.40 GHz) are 3,770 and 22,028. In the Ubuntu VM benchmarks are 3,227 and 13,884. Wow, a huge hit on the IOPs.

I installed 32 bit Ubuntu in the VM along with XFCE desktop because that combination seemed to make the system more stable and responsive. Does 32 bit BOINC benchmark lower than 64 bit BOINC?

I'm burning a 32 bit Fedora 17 Live DVD as I post this. I'll report the benchmarks as soon as I get them. I might have to try a 64 bit Fedora 17 VM for comparison.



Hi, im may case stoping intensive apps. in the host did improve Flops from 1450 to 1800 aprox and IOps from 4900 to 5500 aprox, but these numbbers are still far from 'normal' (2100 and 6300 approx).

I guess the loss is the consequence of having Virtul Box, task manager (occasionaly) and System processes consuming CPU. These processes would not be runnig at the same time (or would not consume so much CPU) so the FLOPs would be higgher without the VM.
ID: 258 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
MarkJ
Avatar

Send message
Joined: 27 Jun 12
Posts: 129
Credit: 62,714,553
RAC: 0
Message 259 - Posted: 17 Sep 2012, 9:08:23 UTC
Have you got vt-i enabled in your bios? Did you select it as an option for the vm? From memory it's under acceleration. Older CPUs don't support it but the newer one should. If you are going to use a 64 bit guest os you must have vt-i support otherwise you are limited to a 32 bit guest os.
BOINC blog
ID: 259 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
MarkJ
Avatar

Send message
Joined: 27 Jun 12
Posts: 129
Credit: 62,714,553
RAC: 0
Message 260 - Posted: 17 Sep 2012, 11:35:11 UTC
On my "guest" OS (Linux Mint 13, 32 bit) which has VT-i enabled I get
3340 whetstone (floating point)
8742 dhrystome (integer)

The "host" OS (Win7 64 bit) gets
4025 whetstone
13094 dhrystone

And these were run without suspending the other one. Some of the difference could be attributed to 64 bit versus 32 bit.
BOINC blog
ID: 260 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
candido

Send message
Joined: 28 Jul 12
Posts: 21
Credit: 2,513,800
RAC: 0
Message 261 - Posted: 17 Sep 2012, 11:37:37 UTC - in response to Message 259.  

Last modified: 17 Sep 2012, 11:38:44 UTC
Hi MarkJ. Thanks for your reply.
I am not an expert, but I think you are asking whether I enabled the virtualization hardware in the BIOS. I checked and my BIOS has nothing related to virtualization. So the option im mention in acceleration is not available in this computer) and I had to install ubuntu 32 bits in a 64 bit windows 7 Host.
Would this then explain the loss in the CPU Flops and IOps?
Thanks
ccandido

EDIT: sorry I posted before seeing your 2nd post.
ID: 261 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
candido

Send message
Joined: 28 Jul 12
Posts: 21
Credit: 2,513,800
RAC: 0
Message 262 - Posted: 17 Sep 2012, 11:42:34 UTC - in response to Message 260.  
I see. Thanks MarkJ
May I ask why you use a 32 bit linux version on this 64 bit computer, especially if this makes you loose some cpu capacity?
I am instaling debian 6.0.5 on another machine and in this case i choose a 64 bit debian. I only use this for runing boinc / asteriods . Is this the best choice?
ccandido

ID: 262 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
jujube

Send message
Joined: 5 Sep 12
Posts: 30
Credit: 24,320
RAC: 0
Message 265 - Posted: 18 Sep 2012, 6:49:13 UTC
So far I've tried a Fedora 17 32 bit VM and a Fedora 17 64 bit VM and both gave very poor BOINC benchmarks. I shutdown all other CPU intensive apps while the benchmarks ran and there was little difference between benchmarks from 32 bit and benchmarks from 64 bit.

Next I'll try a Debian VM and a Gentoo.
ID: 265 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
MarkJ
Avatar

Send message
Joined: 27 Jun 12
Posts: 129
Credit: 62,714,553
RAC: 0
Message 266 - Posted: 18 Sep 2012, 10:21:16 UTC - in response to Message 262.  
I see. Thanks MarkJ
May I ask why you use a 32 bit linux version on this 64 bit computer, especially if this makes you loose some cpu capacity?
I am instaling debian 6.0.5 on another machine and in this case i choose a 64 bit debian. I only use this for runing boinc / asteriods . Is this the best choice?
ccandido


When I was experimenting with Linux mint I was using a Laptop that doesn't have VT-i support so that limited me to a 32 bit OS. Seeing as I had an install disk I used the same on the number cruncher which does have VT-i support.
BOINC blog
ID: 266 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
candido

Send message
Joined: 28 Jul 12
Posts: 21
Credit: 2,513,800
RAC: 0
Message 267 - Posted: 18 Sep 2012, 21:00:39 UTC

Last modified: 18 Sep 2012, 21:22:34 UTC
Hi. The 64 bit Debian 6.0.5 is now running on a i7 2630 QM (laptop) and it has low benchmarks either.
I have exported the other 32 bits Ubuntu VM from my other laptop to this i7 2630 QM.
Took me 20 minutes give or take to export the VM from one laptop to the other!!!!
This means I now have two VM in my i7 laptop.
I ran, at the same time, 4 WU in the 64bit Debian VM and 1 WU in the 32bit Ubuntu VM.
WU in the Debian completed in an average of 36532 secs (10.1 hours) approx, the WU in the Ubuntu took secs 37855 (10.5 hours, +3.6%) (CPU time, not Run time).
I only compared 4 WU in one OS with 1 WU in the other OS, this is not a fair comparison.
But I am keeping the Debian and stoping the ubuntu, although the Boinc Manager is the ubuntu looks beter and the BM in the Debian is the old 6.10.58.

Benchmarks for the debian with activity suspended on the host:

Benchmark results:
Number of CPUs: 4
1273 floating point MIPS (Whetstone) per CPU
11392 integer MIPS (Dhrystone) per CPU

the FLOPS are lower than the Host (1888 FP MIPS)
the integer ops are much higher than the host (5912 I MIPS) !!!!????
ID: 267 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
candido

Send message
Joined: 28 Jul 12
Posts: 21
Credit: 2,513,800
RAC: 0
Message 268 - Posted: 18 Sep 2012, 23:28:27 UTC - in response to Message 265.  

Last modified: 18 Sep 2012, 23:29:41 UTC
hi jujube. Can you quantify the difference?
Although you found a small diference, which one was the better?
ID: 268 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
jujube

Send message
Joined: 5 Sep 12
Posts: 30
Credit: 24,320
RAC: 0
Message 269 - Posted: 19 Sep 2012, 9:45:05 UTC - in response to Message 268.  
hi jujube. Can you quantify the difference?
Although you found a small diference, which one was the better?


I didn't write the numbers down and now I don't remember. The reason I did not write them down is because they were no worse/better than the benchmarks I obtained on Ubuntu.

Tonight I tested with Debian 32 bit and 64 bit. The 32 bit gave 3,023 and 10,880 but the 64 bit gave 3,640 and 22,513 which is very near the 3,770 and 22,028 benchmarks from the host OS. For me 64 bit Debian is the clear winner so I think I will phase out my Ubuntu VMs and replace them with 64 bit Debian VMs.

Ccandido, do I understand correctly that 64 bit Debian did not give good benchmarks for you?

In all cases the host OS on my system is Linux not Windows which some of you may have noticed and which might make you wonder why I bother with these tests since I can run the Asteroid app directly with no need for a VM. One of the reasons I test is because in order to run more than 1 T4T project task at a time one needs to use VMs. The other reason is that all BOINC projects draw their water (CPU cycles) from the same well so it is in the best interest of each project to keep the well (the hosts owned by us crunchers) clean, healthy and efficient, for the good of all parties. Maybe these tests help only a little bit but crunchers know better than anyone how every little bit counts.

Debian repositories seem to make only BOINC 6.10.58 available which might be adequate for Asteroids. If a 7.x version is required then someone in the community will compile a 7.x and make it available or else I will publish instructions for compiling your own. It's easy.
ID: 269 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
candido

Send message
Joined: 28 Jul 12
Posts: 21
Credit: 2,513,800
RAC: 0
Message 271 - Posted: 19 Sep 2012, 19:50:11 UTC - in response to Message 269.  

Last modified: 19 Sep 2012, 20:03:44 UTC

Ccandido, do I understand correctly that 64 bit Debian did not give good benchmarks for you?


Hi jujube, no it did not. the Debian 64 was better than the ubuntu 32 but both these guest OS had worst benchmarks than the host. Much worst!


Debian repositories seem to make only BOINC 6.10.58 available which might be adequate for Asteroids. If a 7.x version is required then someone in the community will compile a 7.x and make it available or else I will publish instructions for compiling your own. It's easy.


At the moment, I am only running asteroids so 6.10.58 is OK, at least for me.
But if you want to post the instructions, someone (more apt than I am) might put them to good use.
ccandido
ID: 271 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile BilBg
Avatar

Send message
Joined: 19 Jun 12
Posts: 221
Credit: 623,640
RAC: 0
Message 296 - Posted: 8 Oct 2012, 7:49:19 UTC - in response to Message 271.  

Last modified: 8 Oct 2012, 7:57:34 UTC

Don't pay so much attention to BOINC benchmarks, they are not some 'Holy Grail', they (normally) differ on different OSes due to different compilers/libs used.

"Interesting difference on my hardware between operating systems":
http://setiathome.berkeley.edu/forum_thread.php?id=69054

Use real apps (run tasks) to compare.

http://www.boinc-wiki.info/Benchmark





- ALF - "Find out what you don't do well ..... then don't do it!" :)
ID: 296 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Kyong
Project administrator
Project developer
Project tester
Project scientist
Avatar

Send message
Joined: 9 Jun 12
Posts: 584
Credit: 52,667,664
RAC: 0
Message 336 - Posted: 27 Oct 2012, 14:21:43 UTC

Last modified: 27 Oct 2012, 14:22:11 UTC
I have done some BOINC benchmarks on virtual machines to comparison BOINC benchmarks between host os and virtualized os.

Host os:

C2D E8400 (7.0.28 64bit) - Win 7 64bit - 3058 floating points MIPS (Whetstone) per CPU, 9129 integer MIPS (Dhrystone) per CPU
GeForce 8600 GT 256 MB - 114 GFLOPS peak

Virtualized:

Windows Server 2012 (7.0.28 64bit) - 2946 floating points MIPS (Whetstone) per CPU, 8921 integer MIPS (Dhrystone) per CPU

Windows XP (7.0.28 32bit) - 3144 floating points MIPS (Whetstone) per CPU, 6304 integer MIPS (Dhrystone) per CPU

Debian 6.0.6 64bit (6.10.58 64bit) - 2904 floating points MIPS (Whetstone) per CPU, 15214 integer MIPS (Dhrystone) per CPU

OpenSuse 12.1 64bit (7.0.28 64bit) - 2968 floating points MIPS (Whetstone) per CPU, 15457 integer MIPS (Dhrystone) per CPU

Ubuntu 12.10 12.1 64bit (7.0.27 32bit) - 2927 floating points MIPS (Whetstone) per CPU, 15978 integer MIPS (Dhrystone) per CPU
ID: 336 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
skgiven
Avatar

Send message
Joined: 28 Jul 12
Posts: 13
Credit: 1,616,360
RAC: 0
Message 377 - Posted: 10 Nov 2012, 9:52:13 UTC - in response to Message 336.  
It's been known for years that Boinc benchmarks vary across different operating systems. Remember that it's just a benchmark, and doesn't necessarily reflect performance. Go by the return times of like for like tasks, not the benchmark. Unfortunately some credit systems are based on the benchmark but ultimately its about returning completed work.

Benchmarks can fluctuate from processor to processor and usually vary between readings on the same machine.
If your reading is around half what you would expect then perhaps the CPU has downclocked, 3.4GHz to 1.6GHz for example.
If you are crunching within a VM and on Windows, the Windows bench if taken without the VM running should be higher; the core freq. of higher end CPU's drops in steps of 100MHz when more cores are used. So a stock i7-2600K @3.4GHz with turbo on would only run one or two threads at 3.8GHz, if running 4 threads each will drop by two hundred MHz and it would be 3.4GHz if 8threads were being used. The benchmark is for one core/thread but all based on Boinc CPU % use setting.
Then there is saturation; the CPU struggles to feed 8threads due to contention. For some projects its been demonstrated that using 7threads is more productive than 8. The benchmark couldn't reflect that as its a generic bench and not specific to the project - which begs the question why don't projects have their own benchmark? This would be more accurate as far as the project is concerned and would help some people decide what system to get/use?
.
ID: 377 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote

Message boards : Windows : FLOPs on a guest virtual machine differ from those of a host machine??