Why SCP and DD from an ESX(i) Host are not Good Performance Tools
I am sure many have noticed that whenever you try to scp a file to or from an ESX(i) host the performance is a little slow. Let’s take the following example, here is an scp between two hosts on the same switch utilizing links that are 1GB.
~# scp 1gb_File 10.10.10.10:/vmfs/volumes/datastore_1 [email protected]'s password: 1gb_File 100% 1000MB 24.4MB/s 00:41
Now running a test with netperf from within two VMs on those two sames host, going through the same path (ie: using the same vmnics), we see the following:
~ # /tmp/netserver Starting netserver at port 12865 ~ # /tmp/netperf -H 10.10.10.11 -t TCP_STREAM -l 30 TCP STREAM TEST to 10.10.10.11 Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec 65536 32768 32768 30.00 939.59
So we are capable of utilizing the full 1GB but the hypervisor is limiting the userworlds other than VMs (ie: dd, scp). I do understand that this is not an apples-to-apples comparison (scp is actually reading and writing data to and from the local disks) but we are seeing scp transferring at 195Mb/s while netperf reaches 939Mb/s.
This is by design and makes perfect sense. We want all our resources dedicated to the VMs and not to any of the tools within the hypervisor. There is a pretty good article entitled I/O Analyzer. It’s a VMware appliance that comes pre-installed with a Linux distro and has IOmeter preloaded on it. I might do a post on how to use that tool, as well. :)
- Recover Files from an SD Card Using Linux Utilities
- Can’t Set X520 10GbE Intel NIC to Auto Negotiate on ESX(i)
blog comments powered by Disqus