On 19 February 2011 00:25, Ivan Voras <ivo...@freebsd.org> wrote:

> No such luck here; I've just tried an amd64 machine (8-STABLE from
> today) in a new installation of XenServer 5.6 and while the GENERIC
> kernel works stable enough, the XENHVM kernel produces all kinds of
> timer-related problems, accompanied by messages like:
>
> Feb 18 23:20:03 xbsd kernel: calcru: runtime went backwards from
> 28669021884109 usec to 22622950 usec for pid 11 (idle)

I've tried comparing the performance of GENERIC and XENHVM kernels on
this machine with unixbench and it points to GENERIC being faster in
everything - though I don't know if this is an artifact of bad timer
behaviour (except for one test, more on this later). Here are the
results:

==> unixbench-generic.txt <==
TEST                                    BASELINE     RESULT      INDEX

Dhrystone 2 using register variables    116700.0 13828395.1     1185.0
Double-Precision Whetstone                  55.0     4104.6      746.3
Execl Throughput                            43.0      885.8      206.0
File Copy 1024 bufsize 2000 maxblocks     3960.0   142874.0      360.8
File Copy 256 bufsize 500 maxblocks       1655.0    65855.0      397.9
File Copy 4096 bufsize 8000 maxblocks     5800.0   146858.0      253.2
Pipe Throughput                          12440.0   936158.7      752.5
Pipe-based Context Switching              4000.0    54004.1      135.0
Process Creation                           126.0     1519.2      120.6
Shell Scripts (8 concurrent)                 6.0      343.0      571.7
System Call Overhead                     15000.0   578046.6      385.4

==> unixbench-xenhvm.txt <==
TEST                                    BASELINE     RESULT      INDEX

Dhrystone 2 using register variables    116700.0 13718470.5     1175.5
Double-Precision Whetstone                  55.0   912662.7   165938.7
Execl Throughput                            43.0      750.2      174.5
File Copy 1024 bufsize 2000 maxblocks     3960.0    96273.0      243.1
File Copy 256 bufsize 500 maxblocks       1655.0    79155.0      478.3
File Copy 4096 bufsize 8000 maxblocks     5800.0    91023.0      156.9
Pipe Throughput                          12440.0   872682.9      701.5
Pipe-based Context Switching              4000.0    50348.4      125.9
Process Creation                           126.0     1511.7      120.0
Shell Scripts (8 concurrent)                 6.0      225.9      376.5
System Call Overhead                     15000.0   561000.3      374.0

Only in the whetstone test (and consistently only in this one across
multiple runs) is the test timing very visibly screwed up, which is
obvious if observing test execution: it takes orders of magnitude
longer than it should (hours) and produces order of magnitude "better"
results than it should. It skews the "final score" so its unusable.

Whetstone is FPU-intensive, unique among these tests.

Just for comparison, here are the results on bare hardware, same OS &
base hardware (MBO, CPU, RAM), different drives & number of CPUs:

TEST                                    BASELINE     RESULT      INDEX

Dhrystone 2 using register variables    116700.0 15419836.3     1321.3
Double-Precision Whetstone                  55.0     3566.9      648.5
Execl Throughput                            43.0     2512.4      584.3
Pipe Throughput                          12440.0  1079209.1      867.5
Pipe-based Context Switching              4000.0    94001.1      235.0
Process Creation                           126.0     4752.0      377.1
System Call Overhead                     15000.0   676244.5      450.8

The result of whetstone indicates that even the GENERIC kernel might
have similar timer problems.
_______________________________________________
freebsd-xen@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-xen
To unsubscribe, send any mail to "freebsd-xen-unsubscr...@freebsd.org"

Reply via email to