Slashdot Mirror
Next G5 Multitasks Operating Systems
squiggleslash writes "IBM has big plans for the 970, Apple's so-called "G5". The CPU will support partitioning, similar to IBM's mainframe systems, allowing multiple operating systems to run at the same time on a single CPU. A Mac built around this chip could theoretically run OS X, GNU/Linux, Mac OS 9, and the PowerPC version of Windows NT, all simultaneously and independently."
a ppc version of the nt kernel will run the next xbox.
#!/usr/bin/english
Partitioning is quite old indeed and has been running on some big iron for a while, but this particular details are somewhat different that VMWare/QEMU.
They plan to add partitioning support on the chip level, so there will be no performance penalty (like in VMWare) or need to recompile OS (like in QEMU IIRC).
Robert
Bastard Operator From 193.219.28.162
..the REALLY cool partitioning is the new POWER5 based iSeries and pSeries. Hell, we've been using a Linux LPAR on our iSeries for two years now.
Now it just works better with POWER5. The FSP (Flexible System Processor) that contains the Hypervisor code is just a card that runs an embedded Linux kernel. Plug that in to a IBM "HMC", or a xSeries Xeon box running SuSE Enterprise 8, which boots into Fluxbox. Open a Java-based config utility, and control all your partitions. Do you see a ongoing theme here? That's right folks, IBM trusts Linux enough to stake the reputation of their Big Iron on it.
With POWER5 all the partitioning is transparent to the OSes. WIth out i5/520 I can move RAM and CPU seamlessly without OS reboots. Hell, I can (and do) have my Linux partition specified with just 2/10ths of one of the POWER5's, with a "burst" limit of 8/10ths. You just setup the FSP/Hypervisor with permissions/profiles for the OSes. If it sees that OS/400 needs more CPU and has a higher priority than Linux, it gets it.
However, this is a very cool move for apple. If I could get a Mac that did all that?
Well, yeah. That would own.
Insert profit.
No, this is different than Vmware etc al
They require a host OS to run under. This means you can partition your CPUs so they can run multiple OSes at the same time, nativly. No need for a host OS, just some bitching BIOS.
There are multiple server vendors who already offer the same ability, mainframes have had this for years.
Of course, running 2 or more OSes requires 2 or more times the CPU power in order to get similar performance to a one OS machine.
If you mod me down, I *will* introduce you to my sister!
The real benefit of true multitasking is of course that you don't have to lock it down. The use of multitasking user mode processes would be quite limited if you only could use it to assign locked processor affinities. Likewise, the ideal situation here would be to let both OSes share both CPUs, with only maybe some additions in the idle loop and perharps an arbitrator driver in each OS. Strictly speaking, I guess an arbitrator would not be needed, but wouldn't it be nice if the OSes could auto-schedule processes of different priorities with each other?
IBM sells a product simply called VM. Actually, I guess it's more of a lease option, as it's only available for mainframes (and I used it on an ES/9000...one of the biggest mainframe (read: MVS/CICS) machines around). It's cool in that you could assign separate processors separate copies of the OS, unlike VMware which has a "host" operating system and then various Guests. There's still a bit of low-level software, but for us it was seemless (which, given how much everything associated with this machine cost, had better have been).
... strange times. Full of promises yet to be fulfilled. But as someone else pointed out, now that OS X is essentialy Unix, there would be precious little reason to go back to the "personality" scheme. I rather think they'd bring out some kick-ass server type box running multiple copies of OS X server, if that is in fact what they're trying to do.
Interestingly, this brought to mind the Pink operating system that IBM and Apple were working on way-back-when(tm). The idea, if I remember correctly, was to have a low level OS kernel that could run multiple personalities...they talked about a MacOS personality (back when System 8 was still being developed), OS/2 and probably some flavor of Unix.
I remember being at what I believe was the last Unix Convention at the Javits Center in NYC around '92 or '93 and they (IBM) had a prototype Power box that purported to be running a super super early pre-alpha version of it. The guy standing by it wouldn't let me touch it, and all he said he could do was run a "DIR" on what was supposed to be the OS/2 personality (no Mac one in sight, for the obvious reason there never was one). He also mentioned that there was a second box, but they couldn't get it to boot.
*Sigh*
I was actually under the impression it was just going to be a dual core PPC, but I RTFA off os OSNews.com a couple of days ago and I don't really remember it.
This would be a JVM running on top of hardware instead of using an intermediate OS layer... Kinda like Sun NCs with Java procs.
Trolling using another account since 2005.
An interesting article and commentary about this Power 5 stuff related to Apple.
The G5 is a 970. The 970 is a G5.
The Power5 is not a 970. They scaled down the Power5 to make the 970.
There are no trails. There are no trees out here.
It's a bit sad that with all the comments posted so far, nobody has mentioned xen, and the fact that it is accomplishing the same kind of thing today on x86 machines with operating systems (including linux 2.4 and 2.6, a couple xbsds, and plan 9) that have been ported to it, and will be able to support un-ported operating systems (e.g. windows xxx) once Intel's silvervale or AMD's pacifica technologies (both are CPU extensions that assist virtualization in hardware) become available (probably in 2006).
They require a host OS to run under. This means you can partition your CPUs so they can run multiple OSes at the same time, nativly. No need for a host OS, just some bitching BIOS.
:). Hypervisor/the FSP on IBM boxes are really insanely cool. BIOS doesn't give it enough credit :)
:).
The CPU has to natively support this as well, at least POWER5 does.
And calling it a BIOS is... almost insulting
Of course, running 2 or more OSes requires 2 or more times the CPU power in order to get similar performance to a one OS machine.
This is depending on what machines you are talking about, all around
WIth IBM, the i5 now has "burstable" limits for CPU. Basically you can assign priority levels for your partitions. Say you have a OS/400 partition and a Linux LPAR. OS/400 gets 80% of CPU and has the highest priority level, and Linux gets 20% and has second priority. Then you throw a AIX partition in and give it some of OS/400's CPU, but assign it the lowest priority. You allow OS/400 and Linux to have burstable limits of CPU, and AIX to not.
So you are running a mailserver on the Linux LPAR. When everyone comes in at 8am, the mailserver gets hammered. Hypervisor sees this, and sees that OS/400 is pretty much unused, so Hypervisor gives Linux 60% of OS/400's CPU allocation.
This all happens without any intervention from you, and without any noticeable different to the OS, aside from the fact that it can suddenly magically kill bits faster.
Then at the same time, let's say everyone starts running DB queries. Hypervisor sees this, and sees that OS/400 has a higher priority than Linux, so it takes back some CPU.
Again, this is all without your input.
And with the higher end Capacity ON Demand systems, you can do really crazy shit. Like a box will ship with 4 active CPUs, but 8 actually installed. You can set that box up so that it will activate those extra CPUs if needed. Completely seemless, completely without your input.
Basically this technology is the future. Intel has nothing, and they know it.
And the really cool thing?
Hypervisor runs on the FSP, which runs linux.
It depends on your definition of efficiency.
If each OS is running on a separate CPU, they can use a UP (uniprocessor) configuration eliminating the overhead in SMP. This buys back typically 5-10% CPU resource, depending on the OS's SMP implementation. However, if one of the OS's CPU demands exceed a single processor's capacity, it will max out extending the transactions (even if the other CPU is idle). However, if you have need for strict separation of resource (e.g., you sell a guarantee of a CPU's processing to a customer), this may be the only way to achieve it.
If both OS share both CPUs, there is additional overhead from the SMP effects, so some CPU resource is "wasted" (not processing the transactions). However, both OS can utilize the resources of both CPUs completely. This would be great for a desktop environment; as you switch between one OS and another, the switched away OS likely goes mostly idle (aside from background processes). In a multiple virtual server environment, it would be useful too; a virtual server that gets "Slashdotted" would be able to obtain on-demand addition resources.
In a CPU contention situation where both OS want to utilize all of both CPUs, the VM (virtual machine here) scheduler will allocate the resource evenly, unless it has a prioritization/capping capability. This is really an inefficient situation, since you have two (or more) OS with SMP overhead and no available CPU resource (maximum CPU going to non-productive use).
You also need to remember when running in a shared CPU mode, there is no guaranteed response time. Applications that need near-realtime response will see an increase in latency variability.
Can You Say Linux? I Knew That You Could.
Actually, they scaled down the Power4 to make the 970.
Hopefully, the following helps clear up some things:
A virtual machine (which for some reason Java-ites seem to believe is *ONLY* a concept which applies to java) "virtualizes" aspects of the host machine. In the case of VMware on intel a technique called a monitor is used to run code natively on the processor for maximum speed. The monitor is basically a pseudo-device which accesses memory directly (you'll see it in the dev dir as vmmon). Other aspects of the machine, such as video, audio, keyboard, network, etc must be "virtualized".
With CPU parititioning, running several operating systems becomes more or less trivial since much of the work of handling all of the ins and outs of virtualizing the hardware is handled for you at the lowest possible level.
So, yes, there may be some performance hit when running several OSes in this way, but not quite as much of one as when this is done totally in software.
GJC
Gregory Casamento
## Chief Maintainer for GNUstep
I don't know what you mean by 'only the protege can access subwindows'. In OS X Command-Tab cycles applications and Command-Backquote cycles windows within the current application.
I find mail fast enough. I'm not sure what you want from 'postfiltering' but I find the rules which let me place email from particular people or mailing lists into particular folders, or forward to my gmail account flexible enough to do anything I need.
Search works in preview for me.
Tom
I have discovered a wonderful