WARNING - OLD ARCHIVES

This is an archived copy of the Xen.org mailing list, which we have preserved to ensure that existing links to archives are not broken. The live archive, which contains the latest emails, can be found at http://lists.xen.org/
   
 
 
Xen 
 
Home Products Support Community News
 
   
 

xen-devel

RE: [Xen-devel] open/stat64 syscalls run faster on Xen VM than standard

To: "xuehai zhang" <hai@xxxxxxxxxxxxxxx>, "Anthony Liguori" <aliguori@xxxxxxxxxx>
Subject: RE: [Xen-devel] open/stat64 syscalls run faster on Xen VM than standard Linux
From: "Petersson, Mats" <mats.petersson@xxxxxxx>
Date: Tue, 29 Nov 2005 13:59:47 +0100
Cc: Tim Freeman <tfreeman@xxxxxxxxxxx>, Kate Keahey <keahey@xxxxxxxxxxx>, Xen Mailing List <xen-devel@xxxxxxxxxxxxxxxxxxx>
Delivery-date: Tue, 29 Nov 2005 13:00:25 +0000
Envelope-to: www-data@xxxxxxxxxxxxxxxxxxx
List-help: <mailto:xen-devel-request@lists.xensource.com?subject=help>
List-id: Xen developer discussion <xen-devel.lists.xensource.com>
List-post: <mailto:xen-devel@lists.xensource.com>
List-subscribe: <http://lists.xensource.com/cgi-bin/mailman/listinfo/xen-devel>, <mailto:xen-devel-request@lists.xensource.com?subject=subscribe>
List-unsubscribe: <http://lists.xensource.com/cgi-bin/mailman/listinfo/xen-devel>, <mailto:xen-devel-request@lists.xensource.com?subject=unsubscribe>
Sender: xen-devel-bounces@xxxxxxxxxxxxxxxxxxx
Thread-index: AcX04mKBxPEKqYLmRRWS8TBe3tYSBQAADfZg
Thread-topic: [Xen-devel] open/stat64 syscalls run faster on Xen VM than standard Linux
 

> -----Original Message-----
> From: xen-devel-bounces@xxxxxxxxxxxxxxxxxxx 
> [mailto:xen-devel-bounces@xxxxxxxxxxxxxxxxxxx] On Behalf Of 
> xuehai zhang
> Sent: 29 November 2005 12:40
> To: Anthony Liguori
> Cc: Kate Keahey; Tim Freeman; Xen Mailing List
> Subject: Re: [Xen-devel] open/stat64 syscalls run faster on 
> Xen VM than standard Linux
> 
> Anthony Liguori wrote:
> > This may just be the difference between having the extra level of 
> > block caching from using a loop back device.
> > 
> > Try running the same benchmark on a domain that uses an actual 
> > partition.  While the syscalls may appear to be faster, I 
> imagine it's 
> > because the cost of pulling in a block has already been 
> payed so the 
> > overall workload is unaffected.
> 
> I created a new domainU by using the physical partition 
> instead of the loopback file as the backends of the VBDs and 
> I reran the "strace -c /bin/sh -c /bin/echo foo" benchmark 
> inside of the domU. The following are the results. Comparing 
> with the results for domU with loopback files as VBDs I 
> reported in the previous email (quoted below), the average 
> time of open/stat64 syscalls are very similar, but still much 
> smaller than the values for standard Linux. If the reason 
> that open/stat64 run faster on domU with loopback files as 
> VBDs is because of the extra level of block caching from 
> using a loop back device, why open/stat64 still run 
> similarily faster on domU with physical partition as VBDs 
> when there is no extra level of block caching from using a 
> loop back device?

Again, have you run the same test-case on the same physical partition?
You should be able to do the same trick with mount and chroot for this
one too. 

This is just to rule out other variations (where in the file-system
different files live, just as an example). 

I'm also not entirely convinced that you don't get some double-caching
with this method too - there is the real hard-disk driver in Dom0, then
the virtual disk driver in DomU, and the file-system driver in DomU. I
don't know if the real or virtual block-device driver itself will cache
things or not. Someone who knows may want to comment on this... 

Still, I agree with Anthony - run some REAL loads on the system, rather
than microbenchmark a particular, very small function of the whole
system. Microbenchmarks are very good at showing how long some
particular thing takes, but they are also very sensitive to extremely
small changes in the system architecture, which for a bigger load would
not be noticable. 

I'm doing my best to explain things here, but I'm not sure if I'm
getting the point across, as it seems like you're trying very hard to
get back to the same original question. We do know that Xen behaves
different than Linux. One reason for this is that it's dealing with
hardware devices through a virtual device. This will most of the time
slow things down, but as you've seen, for example due to the
double-caching effect, you may sometimes actually gain from it,
particularly with very small workloads. 

--
Mats
> 
> XenLinux (physical partition as VBDs)
> 
> root@cctest1:~/c# strace -c /bin/sh -c /bin/echo foo
> % time     seconds  usecs/call     calls    errors syscall
> ------ ----------- ----------- --------- --------- ----------------
>   39.56    0.001955        1955         1           write
>   18.94    0.000936         936         1           execve
>    7.65    0.000378          24        16           old_mmap
>    7.57    0.000374          42         9         2 open
>    6.27    0.000310          52         6           read
>    5.10    0.000252          84         3           munmap
>    4.92    0.000243           9        26           brk
>    1.92    0.000095          14         7           close
>    1.78    0.000088           8        11           rt_sigaction
>    1.40    0.000069          10         7           fstat64
>    1.01    0.000050           8         6           rt_sigprocmask
>    0.93    0.000046          23         2           access
>    0.79    0.000039          13         3           uname
>    0.69    0.000034          17         2           stat64
>    0.38    0.000019          19         1           ioctl
>    0.16    0.000008           8         1           getppid
>    0.16    0.000008           8         1           getpgrp
>    0.14    0.000007           7         1           time
>    0.14    0.000007           7         1           getuid32
>    0.14    0.000007           7         1           getgid32
>    0.12    0.000006           6         1           getpid
>    0.12    0.000006           6         1           getegid32
>    0.10    0.000005           5         1           geteuid32
> ------ ----------- ----------- --------- --------- ----------------
> 100.00    0.004942                   109         2 total
> 
> Thanks.
> 
> Xuehai
> 
> 
> > xuehai zhang wrote:
> > 
> >> Dear all,
> >> When I debugged the execution performance of an application using 
> >> strace, I found there are some system calls like open and stat64 
> >> which run faster on XenLinux than the standard Linux. The 
> following 
> >> is the output of running "strace -c /bin/sh -c /bin/echo 
> foo" on both 
> >> systems. An open call runs averagely 109 usec on standard 
> Linux but 
> >> only 41 usecs on XenLinux. An stat64 call runs
> >> 75 usecs on standard Linux but only 19 usecs on XenLinux.
> >> The Xen VM runs on the same physical machine as the 
> standard Linux. 
> >> It uses loopback files in dom0 as the backends of VBDs.
> >> Any insight is highly appreciated.
> >> Thanks.
> >> Xuehai
> >>
> >> XenLinux:
> >> # strace -c /bin/sh -c /bin/echo foo
> >>
> >> % time     seconds  usecs/call     calls    errors syscall
> >> ------ ----------- ----------- --------- --------- ----------------
> >>  39.05    0.001972        1972         1           write
> >>  19.35    0.000977         977         1           execve
> >>   7.74    0.000391          24        16           old_mmap
> >>   7.23    0.000365          41         9         2 open
> >>   6.06    0.000306          51         6           read
> >>   5.17    0.000261          10        26           brk
> >>   4.93    0.000249          83         3           munmap
> >>   1.98    0.000100          14         7           close
> >>   1.90    0.000096           9        11           rt_sigaction
> >>   1.52    0.000077          11         7           fstat64
> >>   1.01    0.000051           9         6           rt_sigprocmask
> >>   0.95    0.000048          24         2           access
> >>   0.81    0.000041          14         3           uname
> >>   0.75    0.000038          19         2           stat64
> >>   0.38    0.000019          19         1           ioctl
> >>   0.18    0.000009           9         1           time
> >>   0.18    0.000009           9         1           getppid
> >>   0.16    0.000008           8         1           getpgrp
> >>   0.16    0.000008           8         1           getuid32
> >>   0.14    0.000007           7         1           getgid32
> >>   0.12    0.000006           6         1           getpid
> >>   0.12    0.000006           6         1           geteuid32
> >>   0.12    0.000006           6         1           getegid32
> >> ------ ----------- ----------- --------- --------- ----------------
> >> 100.00    0.005050                   109         2 total
> >>
> >>
> >> Standard Linux:
> >> # strace -c /bin/sh -c /bin/echo foo
> >>
> >> % time     seconds  usecs/call     calls    errors syscall
> >> ------ ----------- ----------- --------- --------- ----------------
> >>  22.90    0.000982         109         9         2 open
> >>  22.85    0.000980         980         1           execve
> >>  10.87    0.000466          29        16           old_mmap
> >>  10.45    0.000448         448         1           write
> >>   7.06    0.000303          51         6           read
> >>   6.67    0.000286          10        30           brk
> >>   3.61    0.000155          78         2           access
> >>   3.50    0.000150          75         2           stat64
> >>   2.91    0.000125          42         3           munmap
> >>   2.24    0.000096          14         7           close
> >>   2.12    0.000091          13         7           fstat64
> >>   1.84    0.000079           7        11           rt_sigaction
> >>   1.03    0.000044           7         6           rt_sigprocmask
> >>   0.72    0.000031          10         3           uname
> >>   0.19    0.000008           8         1           geteuid32
> >>   0.16    0.000007           7         1           time
> >>   0.16    0.000007           7         1           getppid
> >>   0.16    0.000007           7         1           getpgrp
> >>   0.16    0.000007           7         1           getuid32
> >>   0.14    0.000006           6         1           getpid
> >>   0.14    0.000006           6         1           getgid32
> >>   0.12    0.000005           5         1           getegid32
> >> ------ ----------- ----------- --------- --------- ----------------
> >> 100.00    0.004289                   112         2 total
> >>
> >> _______________________________________________
> >> Xen-devel mailing list
> >> Xen-devel@xxxxxxxxxxxxxxxxxxx
> >> http://lists.xensource.com/xen-devel
> >>
> > 
> > 
> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@xxxxxxxxxxxxxxxxxxx
> http://lists.xensource.com/xen-devel
> 
> 


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