software:xen:benchmarking

some benchmarking

Testimonials / What ressources to benchmark and how?

network-i/o: netcat -l -p 2222 >/dev/null & dd if=/dev/zero bs=512M count=10 |netcat 10.0.0.1 2222
disk i/o random: iozone, bonnie++
disk i/o block: time bash -c 'dd if=/dev/zero of=testfile bs=512M count=4; sync;'
ram i/o, pure userspace/cpu-power: kernelunzipping to ram, kernelcompilation in ram, ubench

Testcases

hvm/nas/net: domU is in a file on attached nfs in HVM-mode, no paradrivers
hvmp/nas/net: domU is in a file on attached nfs in HVM-mode, paradrivers
para/nas/net: domU is in a file on attached nfs
para/lvm/san: domU is in a logical volume on SAN
para/file/ram: domU is in a file on a dom0-ram
para/lvm/iscsi:domU is in a logical volume on iscsi
para/xfs/iscsi:domU is in a file on an xfs-filesystem on iscsi
para/lvm/disc: domU is in a logical volume on the local disk of dom0
para/xfs/disc: domU is in a file on an xfs-filesystem on the local disk of dom0
dom0/xfs/disc: tests on the dom0 (using internal sas-discs)
bare/xfs/disc: bare metal linux without xen-kernel

results

	hvm	hvmp	paravirtualized domU							dom0	bare
	nas	nas	nas	lvm	file	lvm	xfs	lvm	xfs	xfs	xfs
	net	net	net	san	ram	iscsi	iscsi	disc	disc	disc	disc
network throughput domU→dom0 in Mb/s	7	84	230	229	231	221	226	230	231	246	344
network throughput dom0→domU in Mb/s	21	122	97	110	96	93	93	98	97	246	344
latency for icmp domU↔dom0 in ms	.2	.1	.1	.1	.1	.1	.1	.1	.1	.03	.01
time tar xjf linux-2.6.19.tar.bz2 in s	17	14	25	18	18	17	15	13	18	16	13
time make -j8 bzImage modules (real in s)	1241	1257	561	547	560	555	556	559	559	593	465
time it takes to write a 2GB blockfile	17	72	40	16	8	42	48	139	65	36	32

notice

OS was SuSE SLES10SP1 amd64 with current patches as of september 2007
kernel extraction and compiling was done in ram of the domU
kernel compilation was done w/ 4 cpus pinned to the domU/dom0 and warm cache, 8GB ram
notice: cant test how tap:aio instead of file behaves since sles10sp1 xen doesnt support it apparently
here i used the disc the domU was running on for testing. Below is a test for a disc without partitiontable/filesystem, just blockdevice handed through.

conclusion: many values dont make much sense. random i/o also hasnt been tested yet. One should benchmark having a specific environment in mind.

some more benchmarking

What was benched here: raw i/o throughput. Just wanted to see how much performance-overhad a xen-domU has compared to the dom0 under it. Bevore testing i filled up on a host a partition on local harddiscs with data, and compared this to filling up a lvm-volume: both values are the same as expected. Commands used for testing now:

# in dom0:
time bash -c 'dd if=/dev/zero of=/dev/system/lv_ch_perftest bs=512M count=16; sync;'
# this writes 8GB of data to the logical volume and syncs it. 

# in domU, after handing the lv_ch_perftest over as xvdb:
time bash -c 'dd if=/dev/zero of=/dev/xvdb bs=512M count=16; sync;'