Infrastructure Software

Hello,

we are evaluating a virtualization scenario using Linux and Xen 3.4.0 (OracleVM) but we are observing and measuring really slow disk-writes.

We started with a fresh Ubuntu 2.6.32-22-server install: 4 cpu, 4 GB RAM, HP ProLiant ML350 G6 Server series hard disk, os fully virtualized.
Noticing it's slowliness while writing on disk, we ran a few benchmarks: one with dd, the postgresql pg_bench benchmark and bonnie++.

The results were really discouraging. Less than 4MB/s for writing to disk:

root@ubuntu:/home/rd# time sh -c "dd if=/dev/zero of=bigfile bs=8k count=1000000 && sync"
1000000+0 records in
1000000+0 records out
8192000000 bytes (8.2 GB) copied, 2352.29 s, 3.5 MB/s

Where the dd benchmark against the raw disk runs in approximately 120MB/s.

Less than 6 (!) transaction per second running concurrrently 200 tx with 15 clients:

tps: 5.833331 (excluding connections establishing)

where the result running the same bm on my physical workstation is:

tps: 1462.109434 (excluding connections establishing)

We have no results for bonnie++ because it didn't even complete in a couple of hours.

So we tried other distributions: Oracle Enterprise Linux para-virtualised, SUSE fully and para-virtualised but results were only slightly better. The best we registered was writing at 20MB/s with dd on SUSE fully vrtualised. Still far from our expectations.

The question is: should we expect much better results (I was expecting at least 50 MB/s from dd, and 8-900 tps from pg_bench) and we are doing something wrong?

Thanks,