Being busy with virtualization and additional security measures, I frequently come in contact with people asking me what the performance impact is. Now, you won't find the performance impact of SELinux here as I have no guests nor hosts that run without SELinux. But I did want to find out what one can do to compare system (and later application) performance, so I decided to take a look at the various benchmark utilities available. In this first post, I'll take a look at sysbench (using 0.4.12, released on March 2009 - unlike what you would think from the looks of the site alone) to compare the performance of my KVM guest versus host.
The obligatory system information: the host is a HP Pavilion dv7 3160eb with an Intel Core i5-430M processor (dual-core with 2 threads per core). Frequency scaling is disabled - the CPU is fixed at 2.13 Ghz. The system has 4Gb of memory (DDR3), the internal hard disks are configured as a software RAID1 and with LVM on top (except for the file system that hosts the virtual guest images, which is a plain software RAID1). The guests run with the boot options given below, meaning 1.5Gb of memory, 2 virtual CPUs of the KVM64 type. The CFLAGS for both are given below as well, together with the expanded set given by gcc \${CFLAGS} -E -v - &1 | grep cc1.
/usr/bin/qemu-kvm -monitor stdio -nographic -gdb tcp::1301
-vnc 127.0.0.1:14
-net nic,model=virtio,macaddr=00:11:22:33:44:b3,vlan=0
-net vde,vlan=0
-drive file=/srv/virt/gentoo/test/pg1.img,if=virtio,cache=none
-k nl-be -m 1536 -cpu kvm64 -smp 2
# For host
CFLAGS="-march=core2 -O2 -pipe"
#CFLAGS="-D_FORTIFY_SOURCE=2 -fno-strict-overflow -march=core2
-fPIE -O2 -fstack-protector-all"
# For guest
CFLAGS="-march=x86-64 -O2 -pipe"
#CFLAGS="-fno-strict-overflow -march=x86-64 -fPIE -O2
-fstack-protector-all"
I am aware that the CFLAGS between the two are not the same (duh), and I know as well that the expansion given above isn't entirely accurate. But still, it gives some idea on the differences.
Now before I go on to the results, please keep in mind that I am not a performance expert, not even a performance experienced or even performance wanna-be experienced person: the more I learn about the inner workings of an operating system such as Linux, the more complex it becomes. And when you throw in additional layers such as virtualization, I'm almost completely lost. In my day-job, some people think they can "prove" the inefficiency of a hypervisor by counting from 1 to 100'000 and adding the numbers, and then take a look at how long this takes. I think this is short-sighted, as this puts load on a system that does not simulate reality. If you really want to do performance measures for particular workloads, you need to run those workloads and not some small script you hacked up. That is why I tend to focus on applications that use workload simulations for infrastructural performance measurements (like HammerDB for performance testing databases). But for this blog post series, I'm first going to start with basic operations and later posts will go into more detail for particular workloads (such as database performance measurements).
Oh, and BTW, when I display figures with a comma (","), that comma means decimal (so "1,00" = "1").
The figures below are numbers that can be interpreted in many ways, and can prove everything. I'll sometimes give my interpretation to it, but don't expect to learn much from it - there are probably much better guides out there for this. The posts are more of a way to describe how sysbench works and what you should take into account when doing performance benchmarks.
So the testing is done using sysbench, which is capable of running CPU, I/O, memory, threading, mutex and MySQL tests. The first run of it that I did was a single-thread run for CPU performance testing.
$ sysbench --test=cpu --cpu-max-prime=20000 run
This test verifies prime numbers by dividing the number with sequentially increasing numbers and verifying that the remainder (modulo calculation) is zero. If it is, then the number is not prime and the calculation goes on to the next number; otherwise, if none have a remainder of 0, then the number is prime. The maximum number that it divides by is calculated by taking the integer part of the square root of the number (so for 17, this is 4). This algorithm is very simple, so you should also take into account that during the compilation of the benchmark, the compiler might already have optimized some of it.
Let's look at the numbers.
Run Stat Host Guest
1.1 total 35,4331 37,0528
e.total 35,4312 36,8917
1.2 total 35,1482 36,1951
e.total 35,1462 36,0405
1.3 total 35,3334 36,4266
e.total 35,3314 36,2640
================================
avg total 35,3049 36,5582
e.total 35,3029 36,3987
med total 35,3334 36,4266
e.total 35,3314 36,2640
On average (I did three runs on each system), the guest took 3,55% more
time to finish the test than the host (total
). If we look at the pure
calculation (so not the remaining overhead of the inner workings -
e.total
) then the guest took 3,10% more time. The median however (the
run that wasn't the fastest nor the slowest of the three) has the guest
taking 3,09% more time (total) and 2,64% more time (e.total).
Let's look at the two-thread results.
Run Stat Host Guest
1.1 total 17,5185 18,0905
e.total 35,0296 36,0217
1.2 total 17,8084 18,1070
e.total 35,6131 36,0518
1.3 total 18,0683 18,0921
e.total 36,1322 36,0194
================================
avg total 17,5185 18,0965
e.total 35,0296 36,0310
med total 17,8084 18,0921
e.total 35,6131 36,0194
With these figures, we notice that the guest average total run time takes 1,67% more time to complete, and the event time only 1,23%. I was personally expecting that the guest would have a higher percentage than previously (gut feeling - never trust it when dealing with complex matter) but was happy to see that the difference wasn't higher. I'm not going to start analyze this in more detail and just go to the next test: fileio.
In case of fileio testing, I assume that the hypervisor will take up more overhead, but keep in mind that you also need to consider the environmental factors: LVM or not, RAID1 or not, mount options, etc. Since I am comparing guests versus hosts here, I should look for a somewhat comparable setup. Hence, I will look for the performance of the host (software raid, LVM, ext4 file system with data=ordered) and the guest (images on software raid, ext4 file system with data=ordered and barrier=0, and LVM in guest).
Furthermore, running a sysbench test suggests a file that is much larger than the available RAM. I'm going to run the tests on a 6Gb file size, but enable O_DIRECT for writes so that some caches (page cache) are not used. This can be done using --file-extra-flags=direct.
As with all I/O-related benchmarks, you need to define which kind of load you want to test with. Are the I/Os sequential (like reading or writing a large file completely) or random? For databases, you are most likely interested in random reads (data, for selects) and sequential writes (into transaction logs). A file server usually has random read/write. In the below test, I'll use a combined random read/write.
$ sysbench --test=fileio --file-total-size=6G prepare
$ sysbench --test=fileio --file-total-size=6G --file-test-mode=rndrw --max-time=300 --max-requests=0 --file-extra-flags=direct run
$ sysbench --test=fileio --file-total-size=6G cleanup
In the output, the throughput seems to be most important:
Operations performed: 4348 Read, 2898 Write, 9216 Other = 16462 Total
Read 67.938Mb Written 45.281Mb Total transferred 113.22Mb (1.8869Mb/sec)
In the above case, the throughput is 1,8869 Mbps. So let's look at the (averaged) results:
Host: 1,8424 Mbps
Guest: 1,5591 Mbps
The above figures (which are an average of 3 runs) tell us that the
guest has a throughput of about 84,75% (so we take about 15% performance
hit on random read/write I/O). Now I used sysbench here for some I/O
validation of guest between host, but other usages apply as well. For
instance, let's look at the impact of data=ordered
versus
data=journal
(taken on the host):
6G, data=ordered, barrier=1: 1,8435 Mbps
6G, data=ordered, barrier=0: 2,1328 Mbps
6G, data=journal, barrier=1: 599,85 Kbps
6G, data=journal, barrier=0: 767,93 Kbps
From the figures, we can see that the data=journal
option slows down
the throughput to a final figure about 30% of the original throughput
(70% decrease!). Also, disabling barriers has a positive impact on
performance, giving about 15% throughput gain. This is also why some
people report performance improvements when switching to LVM, as - as
far as I can tell (but finding a good source on this is difficult) - LVM
by default disables barriers (but does honor the barrier=1
mount
option if you provide it).
That's about it for now - the next post will be about the memory and threads tests within sysbench.