Home » Mailing lists » Devel » RSS controller v2 Test results (lmbench )
| RSS controller v2 Test results (lmbench ) [message #13020] |
Thu, 17 May 2007 17:50  |
Balbir Singh
Messages: 491
Registered: August 2006
|
Senior Member |
|
|
Hi, Pavel/Andrew,
I've run lmbench on RSS controller v2 with the following patches
applied
rss-fix-free-of-active-pages.patch
rss-fix-nodescan.patch
rss-implement-per-container-page-referenced.patch
rss-fix-lru-race
(NOTE: all of these were posted on lkml)
I've used three configurations for testing
1. Container mounted with the RSS controller and the tests started within
a container whose RSS is limited to 256 MB
2. Counter mounted, but no limit set
3. Counter not mounted
(1) is represented by cont256, (2) by contmnt and (3) by contnomnt respectively
in the results.
L M B E N C H 2 . 0 S U M M A R Y
------------------------------------
Basic system parameters
----------------------------------------------------
Host OS Description Mhz
--------- ------------- ----------------------- ----
cont256 Linux 2.6.20- x86_64-linux-gnu 1993
contmnt Linux 2.6.20- x86_64-linux-gnu 1993
contnomnt Linux 2.6.20- x86_64-linux-gnu 1993
Processor, Processes - times in microseconds - smaller is better
------------------------------------------------------------ ----
Host OS Mhz null null open selct sig sig fork exec sh
call I/O stat clos TCP inst hndl proc proc proc
--------- ------------- ---- ---- ---- ---- ---- ----- ---- ---- ---- ---- ----
cont256 Linux 2.6.20- 1993 0.08 0.33 4.31 5.93 9.910 0.23 1.59 152. 559. 5833
contmnt Linux 2.6.20- 1993 0.08 0.35 3.25 5.80 6.422 0.23 1.53 161. 562. 5937
contnomnt Linux 2.6.20- 1993 0.08 0.29 3.18 5.14 11.3 0.23 1.37 159. 570. 5973
Context switching - times in microseconds - smaller is better
------------------------------------------------------------ -
Host OS 2p/0K 2p/16K 2p/64K 8p/16K 8p/64K 16p/16K 16p/64K
ctxsw ctxsw ctxsw ctxsw ctxsw ctxsw ctxsw
--------- ------------- ----- ------ ------ ------ ------ ------- -------
cont256 Linux 2.6.20- 1.760 1.9800 6.6600 3.0100 6.5500 3.12000 6.84000
contmnt Linux 2.6.20- 1.950 1.9900 6.2900 3.6400 6.6800 3.59000 14.8
contnomnt Linux 2.6.20- 1.420 2.5100 6.6400 3.7600 6.5300 3.34000 21.5
*Local* Communication latencies in microseconds - smaller is better
------------------------------------------------------------ -------
Host OS 2p/0K Pipe AF UDP RPC/ TCP RPC/ TCP
ctxsw UNIX UDP TCP conn
--------- ------------- ----- ----- ---- ----- ----- ----- ----- ----
cont256 Linux 2.6.20- 1.760 18.9 46.5 19.2 22.9 23.0 28.0 40.0
contmnt Linux 2.6.20- 1.950 20.0 44.6 19.2 20.1 37.9 25.2 42.6
contnomnt Linux 2.6.20- 1.420 23.2 38.5 19.2 23.2 24.4 28.9 54.3
File & VM system latencies in microseconds - smaller is better
------------------------------------------------------------ --
Host OS 0K File 10K File Mmap Prot Page
Create Delete Create Delete Latency Fault Fault
--------- ------------- ------ ------ ------ ------ ------- ----- -----
cont256 Linux 2.6.20- 17.6 15.4 62.8 29.4 1010.0 0.401 3.00000
contmnt Linux 2.6.20- 20.7 16.4 68.1 31.9 3886.0 0.495 3.00000
contnomnt Linux 2.6.20- 21.1 16.8 69.3 31.6 4383.0 0.443 2.00000
*Local* Communication bandwidths in MB/s - bigger is better
-----------------------------------------------------------
Host OS Pipe AF TCP File Mmap Bcopy Bcopy Mem Mem
UNIX reread reread (libc) (hand) read write
--------- ------------- ---- ---- ---- ------ ------ ------ ------ ---- -----
cont256 Linux 2.6.20- 382. 802. 869. 1259.5 1757.8 1184.8 898.4 1875 1497.
contmnt Linux 2.6.20- 307. 850. 810. 1236.2 1758.8 1173.2 890.9 2636 1469.
contnomnt Linux 2.6.20- 403. 980. 875. 1236.8 2531.7 912.0 1141.7 2636 1229.
Memory latencies in nanoseconds - smaller is better
(WARNING - may not be correct, check graphs)
---------------------------------------------------
Host OS Mhz L1 $ L2 $ Main mem Guesses
--------- ------------- ---- ----- ------ -------- -------
cont256 Linux 2.6.20- 1993 1.506 6.0260 63.8
contmnt Linux 2.6.20- 1993 1.506 6.0380 64.0
contnomnt Linux 2.6.20- 1993 1.506 6.9410 97.4
Quick interpretation of results
1. contmnt and cont256 are comparable in performance
2. contnomnt showed degraded performance compared to contmnt
A meaningful container size does not hamper performance. I am in the process
of getting more results (with varying container sizes). Please let me know
what you think of the results? Would you like to see different benchmarks/
tests/configuration results?
Any feedback, suggestions to move this work forward towards identifying
and correcting bottlenecks or to help improve it is highly appreciated.
--
Warm Regards,
Balbir Singh
Linux Technology Center
IBM, ISTL
|
|
|
|
|
|
| Re: RSS controller v2 Test results (lmbench ) [message #13036 is a reply to message #13022] |
Fri, 18 May 2007 03:46  |
Balbir Singh
Messages: 491
Registered: August 2006
|
Senior Member |
|
|
Andrew Morton wrote:
> On Thu, 17 May 2007 23:20:12 +0530
> Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
>
>> A meaningful container size does not hamper performance. I am in the process
>> of getting more results (with varying container sizes). Please let me know
>> what you think of the results? Would you like to see different benchmarks/
>> tests/configuration results?
>>
>> Any feedback, suggestions to move this work forward towards identifying
>> and correcting bottlenecks or to help improve it is highly appreciated.
>
> <wakes up>
>
> Memory reclaim tends not to consume much CPU. Because in steady state it
> tends to be the case that the memory reclaim rate (and hopefully the
> scanning rate) is equal to the disk IO rate.
>
With the memory controller, I suspect memory reclaim will become a function
of the memory the container tries to touch that lies outside its limit.
If a container requires 512 MB of memory and we configure the container
size as 256 MB, then we might see aggressive memory reclaim. We do provide
some statistics to help the user figure out if the reclaim is aggressive,
we'll try and add more statistics.
> Often the most successful way to identify performance problems in there is
> by careful code inspection followed by development of exploits.
>
> Is this RSS controller built on Paul's stuff, or is it standalone?
>
It's built on top of the containers infrastructure. Version 2 was posted
on top of containers v8.
> Where do we stand on all of this now anyway? I was thinking of getting Paul's
> changes into -mm soon, see what sort of calamities that brings about.
>
The RSS controller was posted by Pavel based on some initial patches by me,
so we are in agreement w.r.t approach to memory control. Vaidy is working
on a page cache controller, we are able to use the existing RSS infrastructure
for writing the page cache controller (unmapped). All the stake holders are
on cc, I would request them to speak out on the issues and help build a way
to take this forward.
I've been reviewing and testing Paul's containers v9 patches. As and when
I find more issues, I plan to send out fixes. It'll be good to have the
containers infrastructure in -mm, so that we can start posting controllers
against them for review and acceptance.
--
Warm Regards,
Balbir Singh
Linux Technology Center
IBM, ISTL
|
|
|
|
|
|
| Re: RSS controller v2 Test results (lmbench ) [message #13038 is a reply to message #13037] |
Fri, 18 May 2007 04:07 |
Balbir Singh
Messages: 491
Registered: August 2006
|
Senior Member |
|
|
Rik van Riel wrote:
> Balbir Singh wrote:
>
>> A meaningful container size does not hamper performance. I am in the
>> process
>> of getting more results (with varying container sizes). Please let me
>> know
>> what you think of the results? Would you like to see different
>> benchmarks/
>> tests/configuration results?
>
> AIM7 results might be interesting, especially when run to crossover.
>
I'll try and get hold of AIM7, I have some AIM9 results (please
see the attachment, since the results overflow 80 columns, I've
attached them).
> OTOH, AIM7 can make the current VM explode spectacularly :)
>
> I saw it swap out 1.4GB of memory in one run, on my 2GB memory test
> system. That's right, it swapped out almost 75% of memory.
>
This would make a good test case for the RSS and the unmapped page
cache controller. Thanks for bringing it to my attention.
> Presumably all the AIM7 processes got stuck in the pageout code
> simultaneously and all decided they needed to swap some pages out.
> However, the shell got stuck too so I could not get sysrq output
> on time.
>
oops! I wonder if AIM7 creates too many processes and exhausts all
memory. I've seen a case where during an upgrade of my tetex on my
laptop, the setup process failed and continued to fork processes
filling up 4GB of swap.
> I am trying out a little VM patch to fix that now, carefully watching
> vmstat output. Should be fun...
>
VM debugging is always fun!
--
Warm Regards,
Balbir Singh
Linux Technology Center
IBM, ISTL
------------------------------------------------------------ ------------------------------------------------
Test Test Elapsed Iteration Iteration Operation
Number Name Time (sec) Count Rate (loops/sec) Rate (ops/sec)
------------------------------------------------------------ ------------------------------------------------
1 creat-clo 60.00 8885 148.08333 148083.33 File Creations and Closes/second (256 MB container)
1 creat-clo 60.01 8547 142.42626 142426.26 File Creations and Closes/second (unlimited container)
1 creat-clo 60.01 8632 143.84269 143842.69 File Creations and Closes/second (container not mounted)
2 page_test 60.00 6068 101.13333 171926.67 System Allocations & Pages/second (256 MB container)
2 page_test 60.00 5275 87.91667 149458.33 System Allocations & Pages/second (unlimited container)
2 page_test 60.01 5411 90.16831 153286.12 System Allocations & Pages/second (container not mounted)
3 brk_test 60.01 9151 152.49125 2592351.27 System Memory Allocations/second (256 MB container)
3 brk_test 60.02 7404 123.35888 2097100.97 System Memory Allocations/second (unlimited container)
3 brk_test 60.01 8294 138.21030 2349575.07 System Memory Allocations/second (container not mounted)
4 jmp_test 60.01 983062 16381.63639 16381636.39 Non-local gotos/second (256 MB container)
4 jmp_test 60.00 983084 16384.73333 16384733.33 Non-local gotos/second (unlimited container)
4 jmp_test 60.00 982904 16381.73333 16381733.33 Non-local gotos/second (container not mounted)
5 signal_test 60.01 28013 466.80553 466805.53 Signal Traps/second (256 MB container)
5 signal_test 60.00 28360 472.66667 472666.67 Signal Traps/second (unlimited container)
5 signal_test 60.01 28593 476.47059 476470.59 Signal Traps/second (container not mounted)
6 exec_test 60.02 2596 43.25225 216.26 Program Loads/second (256 MB container)
6 exec_test 60.02 2539 42.30257 211.51 Program Loads/second (unlimited container)
6 exec_test 60.01 2536 42.25962 211.30 Program Loads/second (container not mounted)
7 fork_test 60.01 2118 35.29412 3529.41 Task Creations/second (256 MB container)
7 fork_test 60.03 2130 35.48226 3548.23 Task Creations/second (unlimited container)
7 fork_test 60.01 2130 35.49408 3549.41 Task Creations/second (container not mounted)
8 link_test 60.02 47760 795.73476 50131.29 Link/Unlink Pairs/second (256 MB container)
8 link_test 60.02 48156 802.33256 50546.95 Link/Unlink Pairs/second (unlimited container)
8 link_test 60.00 49778 829.63333 52266.90 Link/Unlink Pairs/second (container not mounted)
------------------------------------------------------------ ------------------------------------------------
|
|
|
|
| Re: RSS controller v2 Test results (lmbench ) [message #13143 is a reply to message #13020] |
Mon, 21 May 2007 14:59 |
William Lee Irwin III
Messages: 20
Registered: April 2007
|
Junior Member |
|
|
On Fri, 2007-05-18 at 09:37 +0530, Balbir Singh wrote:
>> oops! I wonder if AIM7 creates too many processes and exhausts all
>> memory. I've seen a case where during an upgrade of my tetex on my
>> laptop, the setup process failed and continued to fork processes
>> filling up 4GB of swap.
On Mon, May 21, 2007 at 09:53:34AM -0400, Lee Schermerhorn wrote:
> Jumping in late, I just want to note that in our investigations, when
> AIM7 gets into this situation [non-responsive system], it's because all
> cpus are in reclaim, spinning on an anon_vma spin lock. AIM7 forks [10s
> of] thousands of children from a single parent, resultings in thousands
> of vmas on the anon_vma list. shrink_inactive_list() must walk this
> list twice [page_referenced() and try_to_unmap()] under spin_lock for
> each anon page.
I wonder how far out RCU'ing the anon_vma lock is.
On Mon, May 21, 2007 at 09:53:34AM -0400, Lee Schermerhorn wrote:
> [Aside: Just last week, I encountered a similar situation on the
> i_mmap_lock for page cache pages running a 1200 user Oracle/OLTP run on
> a largish ia64 system. Left the system spitting out "soft lockup"
> messages/stack dumps overnight. Still spitting the next day, so I
> decided to reboot.]
> I have a patch that turns the anon_vma lock into a reader/writer lock
> that alleviates the problem somewhat, but with 10s of thousands of vmas
> on the lists, system still can't swap enough memory fast enough to
> recover.
Oh dear. Some algorithmic voodoo like virtually clustered scanning may
be in order in addition to anon_vma lock RCU'ing/etc.
On Mon, May 21, 2007 at 09:53:34AM -0400, Lee Schermerhorn wrote:
> We've run some AIM7 tests with Rik's "split lru list" patch, both with
> and without the anon_vma reader/writer lock patch. We'll be posting
> results later this week. Quick summary: with Rik's patch, AIM
> performance tanks earlier, as the system starts swapping earlier.
> However, system remains responsive to shell input. More into to follow.
I'm not sure where policy comes into this.
-- wli
|
|
|
|
|
|
| Re: RSS controller v2 Test results (lmbench ) [message #13176 is a reply to message #13038] |
Mon, 21 May 2007 13:53 |
Lee Schermerhorn
Messages: 2
Registered: May 2007
|
Junior Member |
|
|
On Fri, 2007-05-18 at 09:37 +0530, Balbir Singh wrote:
> Rik van Riel wrote:
> > Balbir Singh wrote:
> >
> >> A meaningful container size does not hamper performance. I am in the
> >> process
> >> of getting more results (with varying container sizes). Please let me
> >> know
> >> what you think of the results? Would you like to see different
> >> benchmarks/
> >> tests/configuration results?
> >
> > AIM7 results might be interesting, especially when run to crossover.
> >
>
> I'll try and get hold of AIM7, I have some AIM9 results (please
> see the attachment, since the results overflow 80 columns, I've
> attached them).
>
> > OTOH, AIM7 can make the current VM explode spectacularly :)
> >
> > I saw it swap out 1.4GB of memory in one run, on my 2GB memory test
> > system. That's right, it swapped out almost 75% of memory.
> >
>
> This would make a good test case for the RSS and the unmapped page
> cache controller. Thanks for bringing it to my attention.
>
> > Presumably all the AIM7 processes got stuck in the pageout code
> > simultaneously and all decided they needed to swap some pages out.
> > However, the shell got stuck too so I could not get sysrq output
> > on time.
> >
>
> oops! I wonder if AIM7 creates too many processes and exhausts all
> memory. I've seen a case where during an upgrade of my tetex on my
> laptop, the setup process failed and continued to fork processes
> filling up 4GB of swap.
Jumping in late, I just want to note that in our investigations, when
AIM7 gets into this situation [non-responsive system], it's because all
cpus are in reclaim, spinning on an anon_vma spin lock. AIM7 forks [10s
of] thousands of children from a single parent, resultings in thousands
of vmas on the anon_vma list. shrink_inactive_list() must walk this
list twice [page_referenced() and try_to_unmap()] under spin_lock for
each anon page.
[Aside: Just last week, I encountered a similar situation on the
i_mmap_lock for page cache pages running a 1200 user Oracle/OLTP run on
a largish ia64 system. Left the system spitting out "soft lockup"
messages/stack dumps overnight. Still spitting the next day, so I
decided to reboot.]
I have a patch that turns the anon_vma lock into a reader/writer lock
that alleviates the problem somewhat, but with 10s of thousands of vmas
on the lists, system still can't swap enough memory fast enough to
recover.
We've run some AIM7 tests with Rik's "split lru list" patch, both with
and without the anon_vma reader/writer lock patch. We'll be posting
results later this week. Quick summary: with Rik's patch, AIM
performance tanks earlier, as the system starts swapping earlier.
However, system remains responsive to shell input. More into to follow.
>
> > I am trying out a little VM patch to fix that now, carefully watching
> > vmstat output. Should be fun...
> >
>
> VM debugging is always fun!
For some definition thereof...
Lee
|
|
|
|
|
|
|
|
| Re: RSS controller v2 Test results (lmbench ) [message #13279 is a reply to message #13278] |
Thu, 24 May 2007 08:00 |
Balbir Singh
Messages: 491
Registered: August 2006
|
Senior Member |
|
|
Paul Menage wrote:
> On 5/24/07, Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
>> Kirill Korotaev wrote:
>> >> Where do we stand on all of this now anyway? I was thinking of
>> getting Paul's
>> >> changes into -mm soon, see what sort of calamities that brings about.
>> > I think we can merge Paul's patches with *interfaces* and then
>> switch to
>> > developing/reviewing/commiting resource subsytems.
>> > RSS control had good feedback so far from a number of people
>> > and is a first candidate imho.
>> >
>>
>> Yes, I completely agree!
>>
>
> I'm just finishing up the latest version of my container patches -
> hopefully sending them out tomorrow.
>
> Paul
Thats good news! As I understand Kirill wanted to get your patches
in -mm and then get the RSS controller as the first candidate in
that uses the containers interfaces and I completely agree with
that approach.
--
Warm Regards,
Balbir Singh
Linux Technology Center
IBM, ISTL
|
|
|
|
Goto Forum:
Current Time: Sun Apr 28 03:50:56 GMT 2024
Total time taken to generate the page: 0.02608 seconds
|
OpenVZ Forum
Members
Search
Help
Register
Login
Home
