Given the fact the RedHat is not pushing KVM out of the Fedora testing ground and into RHEL 5.4 (more-or-less displacing Xen), I'd consider setting up a benchmark comparing it to Xen and VMWare Server.
Should be interesting.
(Read: In my own somewhat limited experience, KVM's ability to scale nicely into >= 8 cores >4GB RAM puts it as a possible alternative to VMWare ESX*... but as I said, my experience it limited to my own deployments...)
Are you running HTTP servers on these instances? If so, 2.5KB for tcpsendbuf sounds too small. It should be big enough for each socket to hold an entire mean-size HTTP response. That number is very application-specific, but 2.5KB is too small for anything but a toy app.
The 2.5kB is not the amount a single buffer will get assigned to it. It's a value that, in case the tcpsndbuf limit has been reached (which you can set as high as you'd like), is reserved to assure the responsiveness of the connections, however slow.
So, the idea is that you set the limit of tcpsnfbuf as high as you'd prefer, and for the barrier value, increase that amount by at least "numproc * 2.5kB".
This is really cool stuff. Looking forward to it.
(I'm kind of surprised to see such bench. on Anandtech - most bench. here are for windows desktops so this seems like fresh air)
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
6 Comments
Back to Article
gilboa - Monday, January 25, 2010 - link
Given the fact the RedHat is not pushing KVM out of the Fedora testing ground and into RHEL 5.4 (more-or-less displacing Xen), I'd consider setting up a benchmark comparing it to Xen and VMWare Server.Should be interesting.
(Read: In my own somewhat limited experience, KVM's ability to scale nicely into >= 8 cores >4GB RAM puts it as a possible alternative to VMWare ESX*... but as I said, my experience it limited to my own deployments...)
- Gilboa
gilboa - Monday, January 25, 2010 - link
%s/RedHat is not/RedHat is now/gkbob - Friday, January 22, 2010 - link
Are you running HTTP servers on these instances? If so, 2.5KB for tcpsendbuf sounds too small. It should be big enough for each socket to hold an entire mean-size HTTP response. That number is very application-specific, but 2.5KB is too small for anything but a toy app.LizVD - Monday, January 25, 2010 - link
The 2.5kB is not the amount a single buffer will get assigned to it. It's a value that, in case the tcpsndbuf limit has been reached (which you can set as high as you'd like), is reserved to assure the responsiveness of the connections, however slow.So, the idea is that you set the limit of tcpsnfbuf as high as you'd prefer, and for the barrier value, increase that amount by at least "numproc * 2.5kB".
karlkesselman - Friday, January 22, 2010 - link
This is really cool stuff. Looking forward to it.(I'm kind of surprised to see such bench. on Anandtech - most bench. here are for windows desktops so this seems like fresh air)
gwolfman - Friday, January 22, 2010 - link
AnandTech has a whole IT section where to do lots of relevant and useful reviews; ones I haven't seen other sites come close to.Anand knows what he's doing, no worries there.
I am also looking forward to the results.