BETA2 - Abysmal speed on Centos5 guest

[_litz]

BETA2 has severely screwed Centos5 --- they are so slow, as to be unusable.

GA installation under Centos 5 was a 45 minute install process.

This normally takes maybe 2 mins tops.

Whole time, CPU was at a full 100% on the guest, and the process was at 100% on the host.

Installing GA has not affected the speed on the guest at all. It's running at something like 1% of normal usable speed.

Host in question, is an 8 core host, 26gb of memory. Guest was only guest running, and uses 2gb memory.

So there is no possible way it was a host resource issue.

This is 100% virtualbox and its guest, and this is most definitely a behavior change between BETA1 and BETA2.

Please advise what you need doc-wise, and I'll capture/provide it.

[Technologov]

Host1: Windows XP, Intel Core 2 Q6600, VBox 3.2.0-BETA2b
Host2: Mandriva Linux 2010, x64, Intel Core 2 Q6600, VBox 3.2.0-BETA2b

I have all sorts of RHEL (3/4/5, 32+64bit) guests running happily here... quite fast.
Please upload vbox.log so we can see what is different on your setup...

[fixedwheel]

@Technologov, have a look at _litz other thread in "linux guests" => http://forums.virtualbox.org/viewtopic.php?f=3&t=29265

[Technologov]

OK, I have uploaded his VBox.log (from other thread, VBox 3.1.6)

Analyzing...

_litz Host OS: CentOS 5.4, x64
_litz Host CPU: Pentium 4 class Xeon

(easy to identify from Core 2-class CPUs: Core 2 "Supports Supplemental SSE3 = 1" in log)

Those CPUs had very early hardware virtualization available... few virtualization instructions, plus slow execution of those instructions, plus I think VirtualBox is not optimized for old style VT-x.
_litz: Please try to disable hardware virtualization, and tell us if speed improves. NOTE: only 32-bit, 1 vCPU guests will run that way.
Software virtualization might perform better.

More: I have read those 3 pages of the other thread. I am sure we will soon find the answer.

_litz: another question: Do you have another host, preferably with Intel Core i7 CPU around?
_litz: One more thing: your concern that VirtualBox supports Fedora better than RHEL is wrong !
In my experience, they are supported about equally. Both on Host and as a guest. Look here:
http://forums.virtualbox.org/viewtopic.php?f=1&t=25735

[_litz]

I've tried turning off VT-x before; it more hurt than it helped.

My problem with BETA2 is that I can see a marked decrease in performance from 3.1.6 to BETA1, and a monstrous decrease from BETA1 to BETA2.

It took 45 minutes to install Guest Additions on BETA2.

That's -- at most -- a two minute operation.

Unfortunately, I do not have a handy i7 to test with ... the IBM x3650 is the platform I was given to implement virtualization on, so my sandbox is rather small.

The same box, btw, flies running VMware Server 2.0 guests with the same exact OS and software loadout as on virtualbox.

- _litz

[fixedwheel]

The same box, btw, flies running VMware Server 2.0 guests with the same exact OS and software loadout as on virtualbox.

just to sort that out: you don't have installed VMware the same time running your slow Virtualbox guests?

[_litz]

That is correct.

[Technologov]

45 minutes? I presume you mean Windows Additions, right ?

OK, lets make few tests together, so we have very rough numbers to speak.

I have a laptop (!) with RHEL 5, x64 host, so lets try to compile a kernel on it. Both host and guest.
Then compare to your hardware, OK ?

Host: RHEL 5.4, 64-bit, Core 2 Duo P8600 2.4 GHz, 4GB of RAM, VirtualBox 3.2.0-BETA2b r61317
OLD: Guest: RHEL 5.4, 32-bit, 1 CPU, 1 GB of RAM + IO APIC, uses SATA (so we can test both VT-x and software mode)
Guest2: RHEL 5.4, 32-bit, 1 CPU, 1 GB of RAM + IO APIC, uses IDE (so we can test both VT-x and software mode)

We need four raw numbers, and two ratios:
1. compile speed on my host - 3:46 min
2. compile speed on my guest - 5:34 min
3. compile speed on your host
4. compile speed on your guest
A. ratio of performance drop on my hardware (VM vs native) Efficiency: 67% (two thirds of native performance)
B. ratio of performance drop on your hardware (VM vs native) Efficiency: __%

Another test: compiling with "oldconfig"
1.a. compile speed on my host - 16:45 min
2.a. compile speed on my guest - 46:21 min
A. ratio of performance drop on my hardware (VM vs native) Efficiency: 36% (1/3)

hmm... it looks as if the more I/O burst VBox has to eat, the slower it works...

Compiling guest additions:
1:10 min (70 seconds ! yay ! - very far from your 45 min number)

Now our setups are very close, software wise. only different hardware.

compile settings: (untar vanilla kernel 2.6.18 from Linus, here)
--OLD: 1. make oldconfig
--OLD: 1.a. answer "Enter" to all questions... this will choose defaults
1. make defconfig
defconfig compiles very few modules, so this is better for benchmarking...
2. make clean
3. make all

Detailed info: (written by me)
http://chorgan.provo.novell.com/lfl/.ht ... ernel.html

NOTES:
1. I am using hand seconds clock.. (timer/stopper). so VM timing wont affect me.
2. compile speed on my host: (16:45 min with oldconfig)

[_litz]

I would try and compile a kernel ... but under centos, I've never been able to successfully manage it without errors.

It's not exactly for the faint of heart.

The GA install that took 45 minutes was a RH EL5 (centos5) 5.4 x86 guest. Host is EL5/Centos5 5.4 x64.

Our processors are similar, but I have twin 3.0's ... and 26gb memory.

Something perhaps that might be easier would be compiling asterisk.

For the purposes of this test though - just installing the GA works just fine.

3.1.6, takes about 2 minutes. BETA1, takes about 5. BETA2, takes 45. Yes, seriously -- 45.

When something "takes a long time" you see guest CPU shoot to 100%, the core the guest's process is attached to on the host jumps to 100%, and the whole guest just slows to a crawl.

- _litz

[Technologov]

>Our processors are similar, but I have twin 3.0's ... and 26gb memory.

Difference between Core 2 and Pentium 4 architecture is *HUGE*.

nevertheless I have finished compiling kernel on host.

Attached my RHEL5 VBox log for detailed comparison...

[Technologov]

Here are my benchmarks from 2007... I have not done any more recent benchmarks (except for Network)

[Technologov]

hmmm.... my guest looks like stuck... on IO write operation, that never finishes...
yea, situations is bad. After 1 hour, my VM still stuck on write IO operation...
This IO trouble started after 37 minutes of compiling...
will try with IDE now

hmm...
opened bug:
http://forums.virtualbox.org/viewtopic. ... 27#p136827

[Technologov]

compilation finished. numbers updated. look above.

Works fine with IDE.

[Technologov]

Are you still around? Look at the numbers... with IDE it works fine.

[_litz]

That might be a possibility ...

Because of an incompatibility within our company's software (I use these VM's for QA testing), we do not use IDE, only SCSI, SAS or SATA.

And I've seen this problem occur with all of the above.

I can try a test with IDE, but I'd have to build a new vm from scratch just for that, and that's not something I have time to do, at the moment; it will likely be several days before I can do that.

It's worth noting that I copied one of my VB vm's to VMware server 2.0 -- just took the .vdi, converted to .vmdk, and booted it ... the same exact tests I ran under VB that took forever ran almost realtime under vmware.

- _litz