More than 2 Processor Support in Panther?

[ffakr]

Re: Quad PowerMac G5's

Originally posted by jaedreth
1) Why have ethernet cards in a cluster box if it's connection is a fibre channel? [/i]

FC isn't a prefered method of connecting cluster nodes. Most people opt for ethernet (MUCH cheaper, but fast enough for many clusters), or they go for something like Myrinet which is really expensive but the latency is hyper low, like a few nano-seconds.

2) Is Firewire really necessary on a cluster? Each cluster should all be completely controllable from a main XServe server that is decked out with such features.

yes.
Firewire on the xserve has many features. It allow you to setup/rebuild server images with a drive/ipod. It allows you to put an entire xServer into a firewire disk mode so you can copy images onto it. It allows you a VERY CHEAP IP-over-a-chainable-bus... so you can have a 400Mbit backend IP network (usually used by controller to manage end nodes in small clusters)

3) The HD on such a cluster doesn't have to be large at all, and only one would be needed, thanks to XRaid.

There isn't a real advantage to shipping a smaller hard drive since 40gb is really the smallest in production and they don't cost much less than 60gb models. Apple probably more than makes up the cost difference by not having to inventory yet another part.

4) If OS X Server were a mature enough operating system, and could configure clusters simply by plugging them in, including restore, initialize, and install functions over the network, an optical drive would not be needed on the cluster machine.

Apple provide a million and one ways to set up a cluster node without using an optical drive. The ones I can remember quickly: fw target disk mode, config script from ldap server (with dhcp), ipod or other fw drive with config script, simply netbooting an image...
BTW, xserve cluster node DOESN'T ship with an optical drive.

5) Graphics card? Why would a cluster need a display connected?

it doesn't, and you can setup a whole rack of cluster nodes without ever hooking up a monitor, mouse, or keyboard to the xserve. BTW... the cluster node DOESN'T have a video card.

Again, if Apple realized that they should sell clusters assuming that no one in their right mind would use them for anything else *other* than clustering, they could make far more powerful clusters without any unneeded features, and actually make them cheaper than regular XServe boxes.

they ARE cheaper than the regular XServe nodes. The dual proc cluster node is the same price as the single processor xServe. The dual cluster is around $1000 cheaper than the dual xserve.

But again, that's assuming one thing... Apple Intelligence. Infer as you will.

are you sure it's Apple's intelligence you need to assume? Perhaps you need to research the subject more. :wink:

Also, the main reason why XServes are not seeing more migration is that it is not a serious enough Unix server. Apple needs to do what it takes to get OS X Server (server only) certified as POSIX compliant.

Do you know what is required to certify an OS POSIX compliant? The spec is thousands of pages long and it requires an intensive review and a big chunk of change.
Right now, Apple provides one major update per year and many more incremental updates in the same period. If Apple insisted on POSIX compliance, they would be able to provide a minor update ever year and a couple major ones per decade.
A lot of Unices that are considered heavy weights (like the *BSD family) do not, to the best of my knowledge, hold posix compliance... yet they are reviered in IT shops around the world.

Also, Apple needs the option in the OS to use all the wonderful OS X server features,..... I work for a company that makes extensive use of unix servers, and there is no way that XServe or OS X could do what we currently do with our servers. It's not mature enough as a unix operating system. So I can't in all honestly suggest anyone leave their current servers for XServe unless they plan to only use the GUI.

I've snipped a lot of your specifics for brevity. If I understand what you're saying, OS X can currently do all of these things.
OS X can do just about everything out of the box that other Unices can. I'm running Panther Server and this one OS could replace every unix infrastructure box on my current campus and every infrastructure box at my last campus (I work at a major university with a HUGE unix/linux installed base)
Panther server offers, out of the box:

• postfix mail
• mysql database
• dhcp with MAC address tables
• apache w/ php, perl... preconfigured
• JBOSS j2ee application server
• webmail with ssl
• a list server
• Kerberos KDC
• LDAP server (with support for passthru, replication...)
• remote console server (ARD)
• smb, cifs, nfs, atalk file server (with working file locking)
• Active Directory integration, in the box, no schema modifications needed
• .... and probably a ton of stuff I'm forgetting... I'm doing this from memory

Perhaps you should at least refer to this http://www.apple.com/xserve/specs.html for some basic info on apple's server.

 

[ffakr]

Re: Re: Quad PowerMac G5's

x

 

[WM.]

Re: Re: Re: Re: Re: xstation

Originally posted by ffakr
Apple discussed this at WWDC. They said the new G5 uses, on average, much less power(current) than the MD G4. I believe they said half but I don't recall the numbers.

Aha! You've got inside information!

I guess this is always the danger of posting on these kinds of forums--one can read all the publicly available info (the online specs, the developer note, etc.) and assume that they know as much as anyone else, but there's always the possibility of someone with inside info.

Also, now that I think about it, another part of the G5 power supply that would be barely used during normal operation is the fan supply. I mean, the PS has to support all the fans running at full speed all the time, if necessary, even though the vast majority of the time they might be running at 1/4 speed or whatever.

In average use, the system clocks the G5s down when they are not under load. I _believe_ they said the dual 2GH clocks down to around 1.2GHz when taking naps (not going to sleep).

The developer note has this to say:

The following processor states are defined:
Run: The system is running at maximum processing capacity with all processors running at full speed.
Idle: The system is idling; this is the default state. All clocks are running and the system can return to running code within a few nanoseconds. If the system has no work to do, it will be in idle mode.

So I assume that the processors clock down to 1.2 when they're idling?

It is, however, lower powered even when it is running full bore.

As far as 100W/cpu, not even close.

I think I got that from some article or forum post on this site; I guess I could try to find it. It didn't seem like the most credible number I've ever seen, but until your post it was the only one I'd seen about the G5.

The Athlon isn't even that hot, and the massive Itanium2 is only slightly hotter (125watts).
For the 2GHz G5, think south of 50watts max per cpu. Power consumption goes down rapidly as frequency decreases too. When the G5 is running scaled back, you should expect something slightly over 10watts per cpu.

Yeah, "13 W @ 1.2 GHz" (or whatever) seems to float around the forums a lot; I think it came from an IBM document around the Microprocessor Forum.

Thanks for clearing all that up.

WM

 

[ffakr]

Re: Re: Re: Re: Re: Re: xstation

Originally posted by WM.
The developer note has this to say:So I assume that the processors clock down to 1.2 when they're idling?[/i]

that sounds about right. I got the details about clocking down from a second hand perspective. I was listening to an engineer in the G5 performance lab... but that sounds about right.

Thanks for clearing all that up.

NP.

I get the feeling I come across like a prick in a lot of posts. I'm really just a pussycat... but I can't stand FUD, and my other posts are usually pretty dry.
I figured I'd take this opportunity to appologize if I offend anyone.

 

[WM.]

Re: Re: Re: Re: Re: Re: Re: xstation

Originally posted by ffakr
I get the feeling I come across like a prick in a lot of posts. I'm really just a pussycat... but I can't stand FUD, and my other posts are usually pretty dry.
I figured I'd take this opportunity to appologize if I offend anyone.

Yeah, me too--amen to all of that. If you're referring to your response to "jaedreth", I guess I'd like to say the same thing about mine. I just don't get a good vibe from him/her. I think we both picked up on it.

If you're referring to your response to my post(s), well, you certainly didn't come across as a prick to me.

WM

 

[MacRAND]

Clusters of Dual or Quad G5s

Rehash:

Panther is designed to handle multiple CPUs - duals, quads, "n" up to 64

Apple and IBM could be associated to develop and manufacture computers with n processors, where n could go to as high as 64 G5!
The project is internally named "Dark Star".

"N" processor computers beyond Dual or Quad are more likely to be xServe

IBM bought and possesses the technology to cluster almost limitless numbers of CPUs, which it might share with or license to Apple.

New $3-billion IBM plant in NY is currently experiencing under-use of its capacity to handle PPC demands by Apple and others for existing chip designs (PPC) as well as specialized chips made for Nvidia and others.

My concern in expecting Apple to design and manufacturer a larger box for a quad-G5 is that it does not sound economically feasible, especially since considerable engineering has already been put into the design of air flow for cooling individual and dual G5 chips, which Apple tells us is "software contolled" by OS X Panther. (Apparently part of the reason for abandoning OS 9.2.2, which would require a substantial re-write).

At $3000 per high-end box, two duals could equal a $6000 quad with double the RAM slots, hard drives, ports, etc. if connectabilty as a cluster can be solved.

Query: iCluster?

Is Panther designed to handle CLUSTERS of Macs?

What else in terms of hardware or software (Panther / P server) do G5 dual 2GHz boxes need beyond 100/1000 Ethernet, FW800/400, and Fiber Optic connectivity to CLUSTER and to minimize latency problems?

Can G5 Macs be run in a Cluster as easily as being set up in Target FireWire mode?

Wouldn't it make a lot of sense to buy a half dozen or so G5 dual 2+GHz Macs for use during the day by individuals as a workstation (video edit stations; math / science processing in a student lab, etc.), and then form a cluster for heavy rendering or processing massive mathmatical calculations after hours?

Isn't the economy of off-the-shelf duals even with video boards, ports, PCI slots, and hard drives more economically reasonable and feasible as dual-purpose units - if they could be used to form a cluster for cooperative computations when needed?

Wouldn't dual-purpose G5s effectively bridge the gap from pure individual Macs and the high-end xServes with 8 to 64 G5 chips running $12,000 to $50,000?

And wouldn't duality of use of off-the-shelf units be an easier sell for grants instead of having to justify a one time large investment in limited purpose high-end equipment?

Isn't the incorporation of individual CPUs into a cluster more flexible? Like, purchasing
6 G5 dual 2.0 GHz during this fiscal year, and another
6 G5 dual 3.2 GHz units during the following fiscal year, etc., which all work well in an expanding cluster?

Wouldn't a staggered purchase plan over time avoid expensive obsolesence, especially when compared to a large single unit purchase?

So the question posed is - what would it take for Apple to make Panther and existing G5s work together for large clusters?
Is Panther or Panther Server with multiple G5s enough?

If not, why not?

Having read all the posts on this, there are obviously enough MacGeniuses teckies contributing to this thread to figure this out and explaint it to the rest of us MacDummies.

What are the practical possibilities?

 

[Ballresin]

Oct Processors

Actually, the 970 can be produced in quad-core fashion, with dual processors. That way you can actually get 8 processors in there. I don't know if Panther supports the Dual Quad-Core 970 though. IBM originally had the 970 at 8 cores, they cut it down for price.

 

[ffakr]

Re: Oct Processors

Originally posted by Ballresin
Actually, the 970 can be produced in quad-core fashion, with dual processors. That way you can actually get 8 processors in there. I don't know if Panther supports the Dual Quad-Core 970 though. IBM originally had the 970 at 8 cores, they cut it down for price.

I'm curious as to what your sources are.

By all appearances, IBM designed the 970 for Apple. The Altivec looks like it was slapped onto a stripped down Power4 core. 8 core processors are not what Apple is looking for.

From everything I've seen, the 970 was always designed to be a stripped down version of Power4... not a chip that was 4x or more larger than a dual core Power4.

 

[ffakr]

Re: Clusters of Dual or Quad G5s

Originally posted by MacRAND
At $3000 per high-end box, two duals could equal a $6000 quad with double the RAM slots, hard drives, ports, etc. if connectabilty as a cluster can be solved.

It may or may not equal a quad processor box. The main difference between two duals and one quad is the bandwidth and latency. If you do work that can be broken into discrete sections, then clusters are the way to go... but if your work is multi-threaded where the threads depend on each others result, you'd be better off running them on one big box if possible.

Is Panther designed to handle CLUSTERS of Macs?

Panther does not provide out of the box. You're looking for 'flip a switch and we're clustering' (so am I) but it isn't there yet. I've heard 'off the record' that Apple's doing some pretty damn cool stuff in this area but I don't have details.

What else in terms of hardware or software (Panther / P server) do G5 dual 2GHz boxes need beyond 100/1000 Ethernet, FW800/400, and Fiber Optic connectivity to CLUSTER and to minimize latency problems?

Can G5 Macs be run in a Cluster as easily as being set up in Target FireWire mode?

You can build clusters with G5s or xServes with the built in hardware. GigE or even 100BT is often enough of a fabric for a cluster, though there are third party high speed and low latency options for mac clusters (like myrinet), but the cost A LOT)
FW Target Disk mode turns a box into a big hard drive. This isn't what you want for a cluster. It's all about storage and not about processing power.

Wouldn't it make a lot of sense to buy a half dozen or so G5 dual 2+GHz Macs for use during the day by individuals as a workstation (video edit stations; math / science processing in a student lab, etc.), and then form a cluster for heavy rendering or processing massive mathmatical calculations after hours?

it would for a lot of situations, but not for others. If you need a fast connection between the boxes you'd be better off with a rack of 1U boxes on GigE copper.
What you are talking about was done in the past on NeXT boxes but that type of functionality is only useful in certain situations. There are plenty of instances where you want more nodes or where you need this power during the day. When you are looking for high computational densities, you want 1U rack servers or blades... not desktop cases.

Wouldn't dual-purpose G5s effectively bridge the gap from pure individual Macs and the high-end xServes with 8 to 64 G5 chips running $12,000 to $50,000?

Isn't the incorporation of individual CPUs into a cluster more flexible? Like, purchasing
6 G5 dual 2.0 GHz during this fiscal year, and another
6 G5 dual 3.2 GHz units during the following fiscal year, etc., which all work well in an expanding cluster?

Wouldn't a staggered purchase plan over time avoid expensive obsolesence, especially when compared to a large single unit purchase?

few things here...
some things run better on big smp boxes than on clusters... Databases are a good example, bad for clusters, good for big iron.
Also, it is sometimes very beneficial to have a well balanced cluster. You don't want a thread on a fast node waiting for data from a slow node. If this happens a lot, your performance will degrade considerably.

So the question posed is - what would it take for Apple to make Panther and existing G5s work together for large clusters?
Is Panther or Panther Server with multiple G5s enough?

For a good cluster, you need powerful nodes, a robust OS (preferably one that is easy to remotely manage, or manage with a lot of automation). You also need some software architecture to actually handle the clustering. In the old NeXT days, there was an application that would build clusters from NeXT workstations. One box would find the other nodes and spool off threads to them.
Apple doesn't have this... yet. I've been told they are doing cool things along these lines though.
Apple, at this point, is telling users to look to other solutions. Sun's grid solution is available on OS X. Some applications also have built in clustering architectures (usually client/server in nature). GridMathematica and Shake work this way.

 

[MrMacMan]

I hope there is more then 2 processors support.

Clustering would only happen if panther has muti processor support so clustering is impossible without it.

 

[WM.]

Re: Re: Clusters of Dual or Quad G5s

Originally posted by ffakr
Some applications also have built in clustering architectures (usually client/server in nature). GridMathematica and Shake work this way.

And Xcode.

I think.

WM

 

[ffakr]

Originally posted by MrMacman
I hope there is more then 2 processors support.

Clustering would only happen if panther has muti processor support so clustering is impossible without it.

I assume that, by 'multiprocessor support' in panther, you are saying that the kernel must have multiprocessor support (which it has).

You can, however, have a cluster of single processor machines (p4) running kernels that are compiled for single processor operations. The Kernel's job is to run as the shim between processes and the hardware.. it only manages the local hardware though.

Strictly speaking, clustering is more than possible on OS X right now. Apple had a session on clustering at WWDC.
Right now, there is no support for 'out of the box' clustering... that is you can't flip a switch on a bunch of OS X boxes to enable the execution of threads that are spawned from a master node on the newtwork. This may happen one day, but not right now. It would be very cool if Apple could pull it off. You run a multi-threaded application and it spawns a new thread for every 'cluster node' that your machine finds on the network.

There are a lot of clustering options available though. You can run something like Sun's Grid software on OS X. This is a framework for clustering. You can also run a cluster aware application. These applications usually have a client/server architecture where you install the clients on multiple nodes and they contact the master server for work units. Grid mathematica, xCode, Shake work this way.

 

[iEric]

OooooOoOOOoOooooooOoOO....I wished I had an option to put in 3 G4s in mine
But i guess two is good enough.