On 2/22/07, Dave Hansen <[EMAIL PROTECTED]> wrote:
B). Shared Kernel.
[ snip ] A lot of what you find in documentation is simply out-of-date. Just a few: * The issues with stand-alone dump for Linux running from NSS have been sort-of resolved because there are tools now to process a VMDUMP instead, and that can dump the NSS along with the non-shared storage. But as we learned VMDUMP was not really meant for big Linux guests... * My (yes, indeed!) SAVESYS= option never managed to win the heart of those who rule that world so they found something else. There is (or will be) an option to make a Linux virtual machine (with proper CP privileges) issue DEFSYS and SAVESYS under the covers (like GCS does). * The I/O configuration of Linux is now done with the sysfs and friends, and does not need the kernel command line anymore (which was part of the saved NSS). There also is a way to pass extra parameters to the kernel through the IPL command. * The entire ramdisk thing that modern distributions use is a royal PITA as far as footprint is concerned. I have tried to make that also use NSS but could not get that done. So while you carefully save 1 MB of virtual machine in the kernel, immediately after that the ramdisk unpacks in virtual storage and touches 8 MB or so. The proper way would be a kernel with built-in drivers and avoid the ramdisk at startup. On VM very realistic because we don't need that many different drivers (and the size of the shared kernel does not matter so much). XIP is neat and has a lot of potential but the break-even is fragile. In specific cases (like for one particular application in one class of servers) you can make it work (see the redbooks). But the general case is still pretty hard. You need to select very carefully which parts you share and which you don't share, or the cost will be higher than the gain. I have a lot of neat ideas to make this work, but it is not low hanging fruit right now. IMHO the reason why you don't see more recent information about this is that people don't do it. When you don't measure it you don't know what you're missing, and you only see the complications. Back then at IBM we were proud to have the shared kernel in such a way that there were "no serviceable kernel parts inside" and thus there was no kernel rpm package installed. Neat until people start asking that we run shell scripts that issue RPM commands to find the level of the kernel and compare that to the required level. And we could not convince them that all these 100 servers ran the same level and that we set that at IPL time. And similar things for when we want to share application code. Rob -- Rob van der Heij Velocity Software, Inc http://velocitysoftware.com/ ---------------------------------------------------------------------- For LINUX-390 subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: INFO LINUX-390 or visit http://www.marist.edu/htbin/wlvindex?LINUX-390
