domU using linux-2.6.37-xen-next pvops kernel with CONFIG_PARAVIRT_SPINLOCKS disabled results in 150% performance improvement (updated)

domU using linux-2.6.37-xen-next pvops kernel with CONFIG_PARAVIRT_SPINLOCKS disabled results in 150% performance improvement (updated)

[Dante Cinco]

[-- Attachment #1.1: Type: text/plain, Size: 2303 bytes --]

(Sorry, I accidentally sent the previous post before finishing the summary
table)

For a couple of months now, we've been trying to track down the slow I/O
performance in pvops domU. Our system has 16 Fibre Channel devices, all
PCI-passthrough to domU. We were previously using a 2.6.32 (Ubuntu version)
HVM kernel and were getting 511k IOPS. We switched to pvops with Konrad's
xen-pcifront-0.8.2 kernel and were disappointed to see the performance
degrade to 11k IOPS. After disabling some kernel debug options including
KMEMLEAK, the performance jumped to 186k IOPS but still well below what we
were getting with the HVM kernel. We tried disabling spinlock debugging in
the kernel but it actually resulted in a drop in performance to 70k IOPS.

Last week we switched to linux-2.6.37-xen-next and with the same kernel
debug options disabled, the I/O performance was slightly better at 211k
IOPS. We tried disabling spinlock debugging again and saw a similar drop in
performance to 58k IOPS. We searched around for any performance-related
posts regarding pvops and found two references to CONFIG_PARAVIRT_SPINLOCKS
(one from Jeremy and one from Konrad):
http://lists.xensource.com/archives/html/xen-devel/2009-05/msg00660.html
http://lists.xensource.com/archives/html/xen-devel/2010-11/msg01111.html

Both posts recommended (Konrad strongly) enabling PARAVIRT_SPINLOCKS when
running under Xen. Since it's enabled by default, we decided to see what
would happen if we disabled CONFIG_PARAVIRT_SPINLOCKS. With the spinlock
debugging enabled, we were getting 205k IOPS but with spinlock debugging
disabled, the performance leaped to 522k IOPS !!!

I'm assuming that this behavior is unexpected.

Here's a summary of the kernels, config changes and performance (in IOPS):

                      pcifront   linux
                      0.8.2      2.6.37-xen-next
                      pvops      pvops
Spinlock
debugging enabled,     186k       205k
PARAVIRT_SPINLOCKS=y

Spinlock
debugging disabled,     70k        58k
PARAVIRT_SPINLOCKS=y

Spinlock
debugging disabled,    247k       522k
PARAVIRT_SPINLOCKS=n


system: HP Proliant DL380G7
Xen: 4.1-unstable (Thu Oct 14 12:46:29 2010 +0100 22241:3a5755249361)
dom0: 2.6.32.24-pvops-stable (xen/stable)
domU: linux-2.6.37-xen-next (2.6.37-rc5-pvops)

- Dante

[-- Attachment #1.2: Type: text/html, Size: 3785 bytes --]

[-- Attachment #2: Type: text/plain, Size: 138 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Re: domU using linux-2.6.37-xen-next pvops kernel with CONFIG_PARAVIRT_SPINLOCKS disabled results in 150% performance improvement (updated)

[Konrad Rzeszutek Wilk]

On Mon, Dec 20, 2010 at 05:03:13PM -0800, Dante Cinco wrote:
> (Sorry, I accidentally sent the previous post before finishing the summary
> table)
> 
> For a couple of months now, we've been trying to track down the slow I/O
> performance in pvops domU. Our system has 16 Fibre Channel devices, all
> PCI-passthrough to domU. We were previously using a 2.6.32 (Ubuntu version)
> HVM kernel and were getting 511k IOPS. We switched to pvops with Konrad's
> xen-pcifront-0.8.2 kernel and were disappointed to see the performance
> degrade to 11k IOPS. After disabling some kernel debug options including
> KMEMLEAK, the performance jumped to 186k IOPS but still well below what we
> were getting with the HVM kernel. We tried disabling spinlock debugging in
> the kernel but it actually resulted in a drop in performance to 70k IOPS.
> 
> Last week we switched to linux-2.6.37-xen-next and with the same kernel
> debug options disabled, the I/O performance was slightly better at 211k
> IOPS. We tried disabling spinlock debugging again and saw a similar drop in
> performance to 58k IOPS. We searched around for any performance-related
> posts regarding pvops and found two references to CONFIG_PARAVIRT_SPINLOCKS
> (one from Jeremy and one from Konrad):
> http://lists.xensource.com/archives/html/xen-devel/2009-05/msg00660.html
> http://lists.xensource.com/archives/html/xen-devel/2010-11/msg01111.html
> 
> Both posts recommended (Konrad strongly) enabling PARAVIRT_SPINLOCKS when
> running under Xen. Since it's enabled by default, we decided to see what
> would happen if we disabled CONFIG_PARAVIRT_SPINLOCKS. With the spinlock
> debugging enabled, we were getting 205k IOPS but with spinlock debugging
> disabled, the performance leaped to 522k IOPS !!!
> 
> I'm assuming that this behavior is unexpected.

<scratches his head> You got me. I am really happy to find out that you guys
were able to solve this conundrum.

Are the guests contending for the CPUs (so say you have 4 logical CPUs and
you launch two guests, each wanting 4 vCPUs)? How many CPUs do the guests have?
Are the guests pinned to the CPUS? What is the scheduler in the Hypervisor? credit1?
> 
> Here's a summary of the kernels, config changes and performance (in IOPS):
> 
>                       pcifront   linux
>                       0.8.2      2.6.37-xen-next
>                       pvops      pvops
> Spinlock
> debugging enabled,     186k       205k
> PARAVIRT_SPINLOCKS=y
> 
> Spinlock
> debugging disabled,     70k        58k
> PARAVIRT_SPINLOCKS=y
> 
> Spinlock
> debugging disabled,    247k       522k
> PARAVIRT_SPINLOCKS=n

Whoa....  Thank you for the table. My first thought is that: "whoa, PV byte-locking
spinlocks sure sucks", but then I realized that there are some improvements in
2.6.37-xen-next. Like in the vmap flushing code .. 

Re: domU using linux-2.6.37-xen-next pvops kernel with CONFIG_PARAVIRT_SPINLOCKS disabled results in 150% performance improvement (updated)

[Dante Cinco]

[-- Attachment #1.1: Type: text/plain, Size: 3369 bytes --]

On Tue, Dec 21, 2010 at 8:22 AM, Konrad Rzeszutek Wilk <
konrad.wilk@oracle.com> wrote:

> On Mon, Dec 20, 2010 at 05:03:13PM -0800, Dante Cinco wrote:
> > (Sorry, I accidentally sent the previous post before finishing the
> summary
> > table)
> >
> > For a couple of months now, we've been trying to track down the slow I/O
> > performance in pvops domU. Our system has 16 Fibre Channel devices, all
> > PCI-passthrough to domU. We were previously using a 2.6.32 (Ubuntu
> version)
> > HVM kernel and were getting 511k IOPS. We switched to pvops with Konrad's
> > xen-pcifront-0.8.2 kernel and were disappointed to see the performance
> > degrade to 11k IOPS. After disabling some kernel debug options including
> > KMEMLEAK, the performance jumped to 186k IOPS but still well below what
> we
> > were getting with the HVM kernel. We tried disabling spinlock debugging
> in
> > the kernel but it actually resulted in a drop in performance to 70k IOPS.
> >
> > Last week we switched to linux-2.6.37-xen-next and with the same kernel
> > debug options disabled, the I/O performance was slightly better at 211k
> > IOPS. We tried disabling spinlock debugging again and saw a similar drop
> in
> > performance to 58k IOPS. We searched around for any performance-related
> > posts regarding pvops and found two references to
> CONFIG_PARAVIRT_SPINLOCKS
> > (one from Jeremy and one from Konrad):
> > http://lists.xensource.com/archives/html/xen-devel/2009-05/msg00660.html
> > http://lists.xensource.com/archives/html/xen-devel/2010-11/msg01111.html
> >
> > Both posts recommended (Konrad strongly) enabling PARAVIRT_SPINLOCKS when
> > running under Xen. Since it's enabled by default, we decided to see what
> > would happen if we disabled CONFIG_PARAVIRT_SPINLOCKS. With the spinlock
> > debugging enabled, we were getting 205k IOPS but with spinlock debugging
> > disabled, the performance leaped to 522k IOPS !!!
> >
> > I'm assuming that this behavior is unexpected.
>
> <scratches his head> You got me. I am really happy to find out that you
> guys
> were able to solve this conundrum.
>
> Are the guests contending for the CPUs (so say you have 4 logical CPUs and
> you launch two guests, each wanting 4 vCPUs)? How many CPUs do the guests
> have?
> Are the guests pinned to the CPUS? What is the scheduler in the Hypervisor?
> credit1?
> >
>

We only have one guest which we assign 16 VCPUs, each pinned to its
respective PCPU. The system has 24 PCPUs (dual Westmere). Each of the 16
Fibre Channel devices is affinitized to its own CPU.

xen_scheduler          : credit
(XEN) Using scheduler: SMP Credit Scheduler (credit)
xm sched-credit: Weight=256, Cap=0


> > Here's a summary of the kernels, config changes and performance (in
> IOPS):
> >
> >                       pcifront   linux
> >                       0.8.2      2.6.37-xen-next
> >                       pvops      pvops
> > Spinlock
> > debugging enabled,     186k       205k
> > PARAVIRT_SPINLOCKS=y
> >
> > Spinlock
> > debugging disabled,     70k        58k
> > PARAVIRT_SPINLOCKS=y
> >
> > Spinlock
> > debugging disabled,    247k       522k
> > PARAVIRT_SPINLOCKS=n
>
> Whoa....  Thank you for the table. My first thought is that: "whoa, PV
> byte-locking
> spinlocks sure sucks", but then I realized that there are some improvements
> in
> 2.6.37-xen-next. Like in the vmap flushing code ..
>
>

[-- Attachment #1.2: Type: text/html, Size: 5556 bytes --]

[-- Attachment #2: Type: text/plain, Size: 138 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Re: domU using linux-2.6.37-xen-next pvops kernel with CONFIG_PARAVIRT_SPINLOCKS disabled results in 150% performance improvement (updated)

[Jeremy Fitzhardinge]

On 12/20/2010 05:03 PM, Dante Cinco wrote:
> (Sorry, I accidentally sent the previous post before finishing the
> summary table)
>
> For a couple of months now, we've been trying to track down the slow
> I/O performance in pvops domU. Our system has 16 Fibre Channel
> devices, all PCI-passthrough to domU. We were previously using a
> 2.6.32 (Ubuntu version) HVM kernel and were getting 511k IOPS. We
> switched to pvops with Konrad's xen-pcifront-0.8.2 kernel and were
> disappointed to see the performance degrade to 11k IOPS. After
> disabling some kernel debug options including KMEMLEAK, the
> performance jumped to 186k IOPS but still well below what we were
> getting with the HVM kernel. We tried disabling spinlock debugging in
> the kernel but it actually resulted in a drop in performance to 70k IOPS.
>
> Last week we switched to linux-2.6.37-xen-next and with the same
> kernel debug options disabled, the I/O performance was slightly better
> at 211k IOPS. We tried disabling spinlock debugging again and saw a
> similar drop in performance to 58k IOPS. We searched around for any
> performance-related posts regarding pvops and found two references to
> CONFIG_PARAVIRT_SPINLOCKS (one from Jeremy and one from Konrad):
> http://lists.xensource.com/archives/html/xen-devel/2009-05/msg00660.html
> http://lists.xensource.com/archives/html/xen-devel/2010-11/msg01111.html
>
> Both posts recommended (Konrad strongly) enabling PARAVIRT_SPINLOCKS
> when running under Xen. Since it's enabled by default, we decided to
> see what would happen if we disabled CONFIG_PARAVIRT_SPINLOCKS. With
> the spinlock debugging enabled, we were getting 205k IOPS but with
> spinlock debugging disabled, the performance leaped to 522k IOPS !!!
>
> I'm assuming that this behavior is unexpected.

Yeah, that would be one way to put it.

>
> Here's a summary of the kernels, config changes and performance (in
> IOPS):
>
>                       pcifront   linux
>                       0.8.2      2.6.37-xen-next
>                       pvops      pvops
> Spinlock
> debugging enabled,     186k       205k
> PARAVIRT_SPINLOCKS=y
>
> Spinlock
> debugging disabled,     70k        58k
> PARAVIRT_SPINLOCKS=y
>
> Spinlock
> debugging disabled,    247k       522k
> PARAVIRT_SPINLOCKS=n

Fascinating.

Spinlock debugging ends up bypassing all the paths that
PARAVIRT_SPINLOCKS affects, so that's consistent with the problem being
the paravirt locking code.

Basically, there's 3 reasons paravirt spinlocks could slow things down:

   1. the overhead of calling into the pv lock code is costing a lot
      (very hard to imagine how it would cause this degree of slowdown)
   2. you're hitting the spinlock slowpath very often, and end up making
      lots of hypercalls
   3. your system and/or workload gets a very strong benefit from the
      ticket lock's FIFO properties
   4. (something else entirely)

When you're running with PARAVIRT_SPINLOCKS=y, are you getting a lot of
counts on the per-cpu spinlock irqs?

What happens if you raise the "timeout" threshold?  If you have
XEN_DEBUG_FS enabled, you can do that on the fly by writing it to
/sys/kernel/debug/xen/spinlocks/timeout, or adjust TIMEOUT in
arch/x86/xen/spinlock.c.  In theory, if you set it very large it should
have the same effect as just disabling PARAVIRT_SPINLOCKS (except still
using byte locks).  That should help isolate which of the possibilities
above are coming into play.

The other data in /sys/kernel/debug/xen/spinlocks could be helpful in
working out what's going on as well.

Do you know if there are specific spinlocks being particularly pounded
on by your workload?  I'm guessing some specific to your hardware.

Thanks,
    J

RE: Re: domU using linux-2.6.37-xen-next pvops kernel with CONFIG_PARAVIRT_SPINLOCKS disabled results in 150% performance improvement (updated)

[Lin, Ray]

 

-----Original Message-----
From: xen-devel-bounces@lists.xensource.com [mailto:xen-devel-bounces@lists.xensource.com] On Behalf Of Jeremy Fitzhardinge
Sent: Tuesday, December 21, 2010 9:39 AM
To: Dante Cinco
Cc: Xen-devel; Konrad Rzeszutek Wilk
Subject: [Xen-devel] Re: domU using linux-2.6.37-xen-next pvops kernel with CONFIG_PARAVIRT_SPINLOCKS disabled results in 150% performance improvement (updated)

On 12/20/2010 05:03 PM, Dante Cinco wrote:
> (Sorry, I accidentally sent the previous post before finishing the 
> summary table)
>
> For a couple of months now, we've been trying to track down the slow 
> I/O performance in pvops domU. Our system has 16 Fibre Channel 
> devices, all PCI-passthrough to domU. We were previously using a
> 2.6.32 (Ubuntu version) HVM kernel and were getting 511k IOPS. We 
> switched to pvops with Konrad's xen-pcifront-0.8.2 kernel and were 
> disappointed to see the performance degrade to 11k IOPS. After 
> disabling some kernel debug options including KMEMLEAK, the 
> performance jumped to 186k IOPS but still well below what we were 
> getting with the HVM kernel. We tried disabling spinlock debugging in 
> the kernel but it actually resulted in a drop in performance to 70k IOPS.
>
> Last week we switched to linux-2.6.37-xen-next and with the same 
> kernel debug options disabled, the I/O performance was slightly better 
> at 211k IOPS. We tried disabling spinlock debugging again and saw a 
> similar drop in performance to 58k IOPS. We searched around for any 
> performance-related posts regarding pvops and found two references to 
> CONFIG_PARAVIRT_SPINLOCKS (one from Jeremy and one from Konrad):
> http://lists.xensource.com/archives/html/xen-devel/2009-05/msg00660.ht
> ml 
> http://lists.xensource.com/archives/html/xen-devel/2010-11/msg01111.ht
> ml
>
> Both posts recommended (Konrad strongly) enabling PARAVIRT_SPINLOCKS 
> when running under Xen. Since it's enabled by default, we decided to 
> see what would happen if we disabled CONFIG_PARAVIRT_SPINLOCKS. With 
> the spinlock debugging enabled, we were getting 205k IOPS but with 
> spinlock debugging disabled, the performance leaped to 522k IOPS !!!
>
> I'm assuming that this behavior is unexpected.

Yeah, that would be one way to put it.

>
> Here's a summary of the kernels, config changes and performance (in
> IOPS):
>
>                       pcifront   linux
>                       0.8.2      2.6.37-xen-next
>                       pvops      pvops
> Spinlock
> debugging enabled,     186k       205k
> PARAVIRT_SPINLOCKS=y
>
> Spinlock
> debugging disabled,     70k        58k
> PARAVIRT_SPINLOCKS=y
>
> Spinlock
> debugging disabled,    247k       522k
> PARAVIRT_SPINLOCKS=n

Fascinating.

Spinlock debugging ends up bypassing all the paths that PARAVIRT_SPINLOCKS affects, so that's consistent with the problem being the paravirt locking code.

Basically, there's 3 reasons paravirt spinlocks could slow things down:

   1. the overhead of calling into the pv lock code is costing a lot
      (very hard to imagine how it would cause this degree of slowdown)
   2. you're hitting the spinlock slowpath very often, and end up making
      lots of hypercalls
   3. your system and/or workload gets a very strong benefit from the
      ticket lock's FIFO properties
   4. (something else entirely)

When you're running with PARAVIRT_SPINLOCKS=y, are you getting a lot of counts on the per-cpu spinlock irqs?

What happens if you raise the "timeout" threshold?  If you have XEN_DEBUG_FS enabled, you can do that on the fly by writing it to /sys/kernel/debug/xen/spinlocks/timeout, or adjust TIMEOUT in arch/x86/xen/spinlock.c.  In theory, if you set it very large it should have the same effect as just disabling PARAVIRT_SPINLOCKS (except still using byte locks).  That should help isolate which of the possibilities above are coming into play.

The other data in /sys/kernel/debug/xen/spinlocks could be helpful in working out what's going on as well.

Do you know if there are specific spinlocks being particularly pounded on by your workload?  I'm guessing some specific to your hardware.

Based on lock statistics we got, the top 5 lock contentions are in our software stack and the lock contention number are 351375/259924/246715/160125/156188 respectively. (pretty much consistent with top 5 lock contentions when we were running with hvm kernel) We also note that there is kernel option to enable/disable Big Kernel Lock in .37 pv-ops. How does this BKL interact with PARAVIRT_SPINLOCK ? Does disabling BKL affect the performance ?

Thanks,
    J

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel