h a l f b a k e r yWe have a low common denominator: 2
add, search, annotate, link, view, overview, recent, by name, random
news, help, about, links, report a problem
browse anonymously,
or get an account
and write.
register,
|
|
|
CPU Bus Interconnect
Join motherboards together to form a Beowulf-ish machine, but with memory sharing. | |
The Beowulf cluster is a marvelous invention, but its usefulness is limited by the relatively miniscule bandwidth of even the best network interfaces. For applications requiring memory sharing, conventional "big iron" supercomputers are still used. I propose the making of a standard PC motherboard with
card edge connectors on both sides, a socket on one and a header on the other. These would allow several motherboards to share a CPU bus; thus memory sharing and common access to a single set of peripherals, etc. could be implemented. The components would all have to be of the same type, but would be quite cheap.
(?) Dual ported RAM
http://www.cypress....oducts/CY7C056V.pdf Dual port, commercial [kbecker, Oct 04 2004]
(?) Quad ported RAM
http://www.kip.uni-...ions/SRAM-Paper.pdf Download the paper if you want it. It may soon be gone. [kbecker, Oct 04 2004]
[link]
|
|
We are parallel-processing Borg. Resistance is futile... |
|
|
some systems are kind of like this. Check out the latest SGI systems. It's just not available on intel based systems, which is the most common and actually one of the worst architectures available. |
|
|
I like the idea, but how would conflicts be resolved?
For instance, suppose 2 processors want to read memory at the same time. They could take it in turns, but then memory access would slow down the more you connected together. (You could have separate busses for local and external data to reduce the problem). |
|
|
I guess this sort of thing tends towards having packet-based data transfer, which is basically what you are trying to avoid.. |
|
|
// We are parallel-processing Borg. Resistance is futile... // |
|
|
Stuff like this is under development using high bandwidth optical links to remove the network bottlenecks. |
|
|
not baked - SGI and its ilk cannot be considered everyday hardware. What i proposed is a new feature for standard PC motherboards. All it would cost is the price of two card edge connectors and a cheap IC to eliminate the electrical nightmares. |
|
|
so really all this idea is, is a faster way to connect your machines? |
|
|
I've often wondered if its possible to just directly wire the PCI slots of two machines together. perhaps some circutry between then, but as little as possible. If one machine could call IRQ's on the other for transferign data, rates would be amazing. |
|
|
Or, maybe a connector between memory slots? |
|
|
maybe... maybe.. another way of conecting this new "plug n' share" motherboards !: |
|
|
Motherboards shares CPU bus. Maybe we could have a MASTER motherboard with 1 CPU, RAM, HDs, PCI connectors, USB, etc... and then a lot of SLAVE motherboard, with 1 CPU and 1 xxxMB of RAM each. |
|
|
Now the tricky thing... let's say each motherboard has a 32bit bus and 128 MB RAM... motherboards could be connected REALLY in parallel... so if you have 4 motherboards (master included), you would have a pc with a 128bits bus (32x4)... and 512MB RAM. So...let's say we have a 1GHz CPU with 128MB ram... instead of buying a 2.4 GHz CPU, we could buy 7 slave motherboards... and maybe we could have a PC with a 256bit bus. |
|
|
I don't really know what is faster... a pc running at 2.4GHz with a 32bit bus... or a PC running at 1GHz with a 256bit bus. |
|
|
Maybe manufacturers don't think some way of developing a motherboard with.. lets say.. 512bit bus... or 1Kbit bus (well it should also need a CPU with same bandwidth) |
|
|
This kind of hardware would be fast for huge numbers... on the other hand.. the bottle neck would be on linear calculations.... |
|
|
The bottle neck are those interconnects on the printed circuit board and the pins on the RAM chip. Otherwise it would work just fine. I tried it with 8-bit CPUs and multiported RAM. That worked fine, but you couldn't afford 1Gbyte of dual or quad ported RAM (see links). Otherwise the motherboard manufacturers would not do all those patches with "Northbridge" and "Southbridge." It would just be one RAM port each for CPU, Graphics, Storage, PCI bus. |
|
|
[NickHunter] Go for the 1GHz, 256bit bus. In my test it looked like the product Frequency * Bus_Width determined the performance. |
|
|
Some late-80s to early-90s
Macintoshes had a PDS (Processor
Direct) slot, which was exactly this
- a pin-for-pin connection to the
CPU. I read somewhere recently
that at least one person had built a
parallel system out of 16mHz SE/30s. |
|
|
Can I be bothered digging up the
link? Nope. But a croissant for
you. |
|
|
A PCI based system would not be
as fast as what you are proposing,
which in modern terminology
would be called linking the Front
Side Buses if I'm not mistaken.
The PCI bus is well downstream of
there. |
|
|
The cache slot might be a
candidate - if there is one. |
|
|
I have, on and off for several months now, been
considering a computer that would consist of two
(preferably heterogeneous) motherboards and CPUs in
one case, operating together as one computer. I had
been thinking of just using Ethernet to interconnect
them, but PCIe or FSB would likely be better. Ethernet
is easier for a prototype though. |
|
|
Anyway, the two sub-computers would run some kind of
hypervisor thing that would combine them into a
cluster, and then a VM containing the OS you actually
want to use would run on the cluster. The hypervisor
would use machine learning to determine which tasks to
run on which CPU, to optimize to their different
strengths. |
|
|
I don't remember for sure, but I think I planned to
make the PCIe bus switchable between the two
motherboards on a per-card basis (using a device that
goes between the motherboards and the cards), so they
could share video cards and such. And since those would
just be forwarded into the VM anyway, I guess the
hypervisor could just switch each card to whichever
motherboard had the better weighted combination of
less load at the moment and better performance with
that card, and it would just work transparently. (Maybe I
came up with a better system for card sharing, but I
don't remember it if so. Also, I don't think PCIe is
usually hot-pluggable, so the hypervisor would have to
handle that too.) |
|
|
Regarding the links to PDFs about multi-ported RAM,
I was able to get the first one using the Wayback
Machine, but I couldn't find the second one. |
|
|
Also, those are about SRAM. Does multi-ported
DRAM (let alone multi-ported DDR(2/3/4)) exist, apart
from VRAM and CPU registers? |
|
|
Get some Dell PowerEdge 2950's with fibre channel. Dual 64-bit Xeons, up to 24Gb RAM. Hook up the fibre channel links. |
|
|
Forget edge connectors, too much like hard work. |
|
| |