Any paper on how to do a VMware like stripped host?

[ToddAndMargo]

Hi All,

I have a server application coming up where I will be using a 64 bit CentOS 5.x host with four (32 bit) Windows Server 2008's running as Virtual Machines. I have seen discussions on the sites before to the effect that if you strip down your host, you can achieve the same practical performance as you get with a VMware server. Can anyone give me a URL as to how to do this with CentOS 5.x? What to install and what not install?

Many thanks,
-T

[TerryE]

Todd, Linux is a pretty benign environment so if you have already got it running then there is little point in stripping it down. Doing do will have two main benefits (i) freeing up memory in your host. Probably the single biggest thing you can do is to stop your desktop and X systems, and fall back to SSH based administration. (ii) freeing up the daemon overheads. Again stopping the desktop / X, USB and bluetooth will help here, as will swiching to server optimised kernel.

The other big thing that you need to look at is your I/O performance. If you have multiple disks in you system then using LVM to RAID-1 the logical partitions that you store your VDIs in will help as this will double your read I/O capacity. (This also allows you to do non-stop backup). Also Ext4 is a better filesystem for big files like VDIs.

You also want to look at your W2K8 guests and stop any unnecessary services.

[fixedwheel]

server optimised kernel.

what is this and how does it work?

using kernel 2.6 i have benchmarked my old P-III and P4 boxes running debian stock 250Hz CONFIG_PREEMPT_NONE=Y kernel (this non-preemptive is server optimized?) vs. selfcompiled 300Hz CONFIG_PREEMPT=Y CONFIG_PREEMPT_BKL=Y with differences about 1%

[TerryE]

This is really a general Linux tuning discussion, rather then a VBox one, but what the heck. There are differences otherwise the various Distro vendors wouldn't produce server and workstation variant kernels, but like most of these things the differences are very marginal except in high stress use cases. Most workstation builds or systems with genuine realtime devices attached use a preemptive kernel.
(Kernel mode) device drivers respond to any device interrupts effective immediately and steal cycle from the running user mode process slot. As a result of this a waiting process can become computable. With a preemptive kernel, the rescheduling occurs immediately. With a non-preemptive kernel, the rescheduling occurs on a fixed tick. Preemptive kernels give better perceived responsiveness to the interactive user, especially if you have a single core system. The downside is that the context switch rate can go through the roof and the % system time rise. With modern multi-core servers, with only disk and network attached devices, there is little point in having a preemptive kernel.

[HolgerB]

Todd, I´m not really shure what your main motivation is to use CentOS.

Personally (in your case) I would either use a standard "Server distribution" (for my gusto this would be namely Debian because I have some experiences with it) or grab some already "scaled down" Linux-based VM solution.

Debian with running Xen does a good job for running Win VMs. I haven´t tested Win 2k8 but Win 2k3 and XP runs fine.
If you want to save some time and do not like to tinker around a lot then a free VM solution such as VMWare ESXi or Citrix XenServer Express might also be of interest for you.

The later two are really set up quickly and esp. the XenServer comes with some nice Windows Administration frontend.

No offence to the VBox dev team but I think that XenServer or VMWare ESXi are more taylored towards multiple server hosting.
For me VBox has more of a desktop VM solution similar to VMWare Workstation.

As a general rule of thumb from my limited experiences:
A open Linux distribution requires more manual work, more knowledge but is very flexible and you can analyze problems very detailed.

A "closed" and taylored Linux distribution (XenServer / VMWare ESX) requires less work, less knowledge but you are most likely limited to the on-board functionality. Our two XenServer-driven hosts at work do a very nice job though but I would not consider them better than the first VM-server we set up by using Debian Etch and the standard Xen Hypervisor. Most of the activities I do with the XenServer windows frontend would be possible via SSH with little to no additional work required.


@TerryE:

If you have multiple disks in you system then using LVM to RAID-1 the logical partitions that you store your VDIs in will help as this will double your read I/O capacity. (This also allows you to do non-stop backup). Also Ext4 is a better filesystem for big files like VDIs.

Sorry for asking that dumb but how could RAID-1 double I/O capacity ? I´m aware of the LVM option to use multiple LV´s as some sort of "stripeset" but how does this mix up with RAID-1 ?
I´m just curious.

For the rest I fully agree with you. It´s more a general Linux tuning discussion which can be easily mixed with the typical "that´s my favourite distrib"-discussion

Cheers,
Holger

[ToddAndMargo]

Todd, I´m not really shure what your main motivation is to use CentOS.

Also know as Red Hat Enterprise Linux. I am going with Red Hat / CentOS as I am trusting their judgment that they will keep everything "enterprise class". That means everything works perfectly (almost) and is also sorely out of date. I am also a member of the Samba mail list. I get to see all the bugs that don't get fixed and all the agony users of the bleeding edge are going through that Red Hat is protecting me from. That protection is much appreciated -- my Samba server just work. The out-of-date drives me nuts at times, but I require the boiler plate aspect of an Enterprise Level OS when building server that are suppose to run forever with very little fuss. (The Terminal Servers will keep busier than h... for years to come!)

-T

p.s. my solution to the out-of-date is Vbox and Kubuntu 9.04. I use it whenever I need to do something bleeding edge.

[TerryE]

@Holger, I tend to agree with you about Xen being the natural FLOSS competitor to ESX. They've also done some interesting stuff around interoperability with the MS virtualisation standards so you can run MS VMs unchanged under XenServer and in this case often a lot more efficiently than under a native MS kernel. VBox really plays better in the Workstation / Type II server scenario.

Sorry for asking that dumb but how could RAID-1 double I/O capacity ? I´m aware of the LVM option to use multiple LV´s as some sort of "stripeset" but how does this mix up with RAID-1 ?

(See the Wiki RAID who want the technical background). I am talking about RAID-1 not RAID-0. What most modern RAID-1 drivers do is to allow queued reads to be satisfied by both mirrors doubling the effective read transaction rate. OK, writes still need to go to both drives, however reads normally exceed writes by a factor of at least four and reads tend to slightly more random, so RAID-1 can effectively increase the I/O transaction rate by ~80% over the comparable single spindles.

[HolgerB]

@others: Sorry, off-topic and non-VBox related stuff following

@Todd: Ok, I see your point now.
I know about CentOS bein the "open variant" of RHEL

So if you major concern about other Linux distributions is stability then Debian might be worth a try. While Ubuntu, Fedora and other "more modern" distributions are often pushing the edge and implement things which might better not be included into a production system, Debian was always rocksolid on all systems I´ve seen or set up myself. Of course this doesn´t mean that CentOS is a bad base for a VM-server

Your choice...

XenServer on the other hand is free and the base for it (Xen) has been proven to be very solid in production use. It could be worth a try. As already mentioned I have had good experiences with both the commercial XenServer version as well as with the free open Source Xen on Debian Etch.

@TerryE:
OK, I do understand now about Raid-1. Basically one HDD of the RAID1 reads for example the first block of a file which the second HDD reads the second block of a file at the same time. This is a nice approach. Never thought about this. Honestly I must admit that I have no big experiences with RAID beside using a 2x1TB on a FreeNAS system

Is this RAID-1 "parallel read" support also integrated into the Linux software RAID ?

Beside all this Xen vs. VBox vs. ESX stuff I´m very confident that VBox also would make a great server solution with some changes to the frontend and some more functionality to the CLI tools included. iSCSI target support is already nice.

[fixedwheel]

Beside all this Xen vs. VBox vs. ESX stuff

ehhm ... isnt Red Hat advertising their own now Xen and future KVM based virtualisation? - this should be available with Centos too

and for the slightly offtopic, i think you can hardly compare Centos (RHEL) with Debian. Centos repository isnt nearly that big as Debian but if you want or have to run legacy applications like oracle database, or install generic available rpm packages Centos may be a better choice. I feel that the Red Hat Kernel Hackers apply much sophisticated stuff in their kernels while Debian Kernel Maintainers stick to reliability achieved with simplicity and maturity (even in testing they are still at 2.6.26). Anyway, if i want a special kernel i compile my own and that is easy done the debian way. And when i go bleeding edge my choice is debian testing which is mostly bug free (bugs stay in sid, i hope) except maybe some glitches sometimes, anyway a good choice for a desktop system. Apart from this - @Todd - stick with Centos as that is the one you seem to know and there is nothing wrong with it

[ToddAndMargo]

Beside all this Xen vs. VBox vs. ESX stuff

ehhm ... isnt Red Hat advertising their own now Xen and future KVM based virtualisation? - this should be available with Centos too

The Xen kernel messes up too many of my critical device drivers, such as my extended serial ports.