Index
> Main > How big should I expect the data cache to be? |
| Author |
|
|
revolution 26 Jan 2012, 10:57
Hugh Aguilar wrote: How big should I expect the data cache to be? I suppose this is going to vary from one processor to another, but what can I assume as a reasonable size for a typical processor? |
|||
26 Jan 2012, 10:57 |
|
|
AsmGuru62 26 Jan 2012, 11:50
@Hugh Aguilar:
This link has some information on cachces: http://www.agner.org/optimize/ Early optimization case? Are you certain, that "accessing that data a lot" is really what is slowing down your code? Did you measure the speed of the code? |
|||
|
26 Jan 2012, 11:50 |
|
|
edfed 26 Jan 2012, 18:26
revolution wrote: Cache can vary from 0B to more than 12MB. It also depends upon which level of cache you are thinking of: L1, L2 or L3. for the moment. maybe more cache layers, and more cache quantity (of course). what about a L0 cache? L-1, etc...  and 4giga bytes L0 cache . |
|||
|
26 Jan 2012, 18:26 |
|
|
revolution 26 Jan 2012, 18:49
L0 would be the CPU registers?
|
|||
|
26 Jan 2012, 18:49 |
|
|
edfed 26 Jan 2012, 21:07
why not? if everything is register, it can be fast. in other words, you completelly delete all the layers, and concentrate to the more important thing, databus to alu connection. let say, you have a all in one cpu+ram. no registers. just ram. add ram to ram, mov ram to ram, mul ram by ram, etc.
|
|||
|
26 Jan 2012, 21:07 |
|
|
revolution 27 Jan 2012, 01:58
The larger you make a memory structure the slower it will be. 4G of CPU registers is never going to happen because it will always be more efficient to have a smaller memory structure to improve overall performance.
|
|||
|
27 Jan 2012, 01:58 |
|
|
Hugh Aguilar 27 Jan 2012, 04:35
AsmGuru62 wrote: @Hugh Aguilar: No, I didn't measure. It is not really meaningful to measure when interrupts may be occurring, and the OS is preemptive. It may be possible to shut all of that off during the execution of a section of code if I knew how the OS works --- the Menuet crowd may be able to comment on that. In many cases, my whole program is oriented around processing data in a big data structure --- usually a linked list or a tree. All of the nodes are the same type of record. If that record fits in the L1 cache then I'm pretty sure the program is going to be significantly faster than it would be if the record straddles a cache boundary. Although I am writing a program similar to what I described above, this is mostly a general question, because a lot of my programs are like this. I haven't delved into Agner's writings yet, but I should --- a lot of people have said that that is the best place to start in learning about optimization, which is a subject I know very little about. I did read Abrash's two Zen books, but they are obsolete now I'm told. Isn't it true though, that nobody really knows how to optimize code nowadays because the processor manufacturers don't tell anybody how their processors work internally? We have guidelines for optimization, but we don't know how much of an effect if any they have. We can measure, but that is just empirical evidence; it doesn't prove anything. This reminds me of the voodoo rituals that some folks do --- there is a lot of anecdotal evidence to indicate that they work --- but there is no proof and there never will be.  |
|||
|
27 Jan 2012, 04:35 |
|
|
revolution 27 Jan 2012, 05:35
Hugh Aguilar wrote: Isn't it true though, that nobody really knows how to optimize code nowadays because the processor manufacturers don't tell anybody how their processors work internally? We have guidelines for optimization, but we don't know how much of an effect if any they have. We can measure, but that is just empirical evidence; it doesn't prove anything. This reminds me of the voodoo rituals that some folks do --- there is a lot of anecdotal evidence to indicate that they work --- but there is no proof and there never will be. |
|||
|
27 Jan 2012, 05:35 |
|
|
Hugh Aguilar 27 Jan 2012, 08:02
revolution wrote:
The key word there is "understanding" --- at least on my part, this is sorely lacking. There is a huge difference between familiarity and understanding --- and the difference between the smart, the dumb and pragmatic is that they know what this difference is, they don't know, and they don't care. Your phrase "diminishing returns" pretty much hits the nail on the head though. For the most part, there are no returns. Desktop software is given away for free (as in free beer). Software increases the value of the hardware that it is running on, but software has no value in itself.  |
|||
|
27 Jan 2012, 08:02 |
|
< Last Thread | Next Thread > |
Forum Rules:
|