[kvm-ppc-devel] little initial performance measurement

[Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>]

######## looking at the guest with kvmstat and at the host with vmstat ########

guest idle looks very busy - we need MSR[WE] support - good that jerone is working on that already:
kvmstat
|       dcr_exits |           exits |     light_exits |      mmio_exits |remote_tlb_flush |    signal_exits |
|               0 |           23302 |           23250 |               0 |               0 |              80 |
|               0 |           24147 |           24064 |               0 |               0 |              79 |
|               0 |           24421 |           24303 |               0 |               0 |              78 |

vmstat
procs -----------memory---------- ---swap-- -----io---- --system-- ----cpu----
 r  b   swpd   free   buff  cache   si   so    bi    bo   in    cs us sy  id wa
 1  0      0 183292      0  26196    0    0     0     0  130   169  2 99   0  0
 1  0      0 183292      0  26196    0    0     0     0  112   150  2 98   0  0
 1  0      0 183292      0  26196    0    0     0     0  112   151  1 100  0  0


When doing some small IO in the guest we see a bit host userspace activity (4%) - but not much, therefore I do not expect IO as the part that slows us down atm.
	procs -----------memory---------- ---swap-- -----io---- --system-- ----cpu----
	 r  b   swpd   free   buff  cache   si   so    bi    bo   in    cs us sy  id wa
	 1  0      0 182932      0  26196    0    0     0     0  308   159  4 96  0  0
The portion of mmio exits to light exits is still <1%


######## some basic perf tests for virtio ########

Compare speed of guest and host (very simple tests, host measured without running guest)
I/O - dd on nfs
	rm /root/testddfile; time dd if=/dev/urandom of=/root/testddfile bs 48 countP0
ping for basic latency
	bash-3.00# ping 192.168.1.2 -c 4

------Guest------						-----Host-----
DD								DD
500+0 records in						500+0 records in
500+0 records out						500+0 records out
								
real    0m22.552s						real    0m0.947s
user    0m1.600s						user    0m0.004s
sys     0m20.864s						sys     0m0.856s

64 bytes from 192.168.1.2: icmp_seq=0 ttld time\x12.3 ms	64 bytes from 192.168.1.2: icmp_seq=0 ttld time=0.165 ms
64 bytes from 192.168.1.2: icmp_seq=1 ttld time=8.14 ms	64 bytes from 192.168.1.2: icmp_seq=1 ttld time=0.153 ms
64 bytes from 192.168.1.2: icmp_seq=2 ttld time=7.04 ms	64 bytes from 192.168.1.2: icmp_seq=2 ttld time=0.143 ms
64 bytes from 192.168.1.2: icmp_seq=3 ttld time=7.08 ms	64 bytes from 192.168.1.2: icmp_seq=3 ttld time=0.147 ms

=> latency and throughput are very bad atm.
We need to find out where this speed is lost. Possibly this IO emulation is slow for us, but we also have some "waits" in the boot process which are definetly not I/O bound (e.g. TLB trashing ?) and we know that our guest idle is a busy wait for us until we have implemented that code for MSR[WE].

A little Timing check:
sleep 30s vs. real clock 30s
Host:
Guest: TODO

Little userspace app that heavily uses memory:
I also think about a small userpsace app that allocate&uses memory heavily to push our tlb code e.g. to test any improvements we make there.
I have a similar app here and will post it once I have adaptedt it to our needs and tested it.

Next mail will be a patch to extend our kvmstat informations about light exits.

-- 

Grüsse / regards, 
Christian Ehrhardt
IBM Linux Technology Center, Open Virtualization

-------------------------------------------------------------------------
Check out the new SourceForge.net Marketplace.
It's the best place to buy or sell services for
just about anything Open Source.
http://ad.doubleclick.net/clk;164216239;13503038;w?http://sf.net/marketplace
_______________________________________________
kvm-ppc-devel mailing list
kvm-ppc-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/kvm-ppc-devel