In this course, you will learn to design the computer architecture of complex modern microprocessors.
Basic Cache Optimizations
Loading...
来自 Princeton University 的课程
Computer Architecture
282 个评分
从本节课中
Advanced Caches 1
This lecture covers the advanced mechanisms used to improve cache performance.
与讲师见面
David WentzlaffAssociate Professor
Electrical Engineering
Okay. Welcome to today's ELE 475.
So today, we are going to be moving to a little bit different topic. We're going
back to one of the topics that we talked about at the beginning of class in Lecture
three. We're going to talk about caches, and
we're spending two lectures in caches because they are really important to
performance. In the, we'll say, in the 90s,' 80s' and
90s,' a lot of, lot of computer architectural papers and a lot of effort
were spent trying to use more and more transistors, which Moors Law was giving
the computer architect and applying them to either build bigger caches,
Fancier caches, more set associative caches, harder will say, ways to think
about caches all for higher performance. Cough And we'll be continuing to videotape
these lectures and put them online for everybody.
And so let's, let's start of looking what we're doing today.
So, we're going to start off with a litlle bit of review from Lecture three.
We're going to talk about the three Cs of caches.
So, capacity, Conflict, and compulsory misses, we'll be
talking about the basic cache optimization we had talked about in Lecture three,
which are mostly kind of, make your cache bigger, make your cache more highly
associative somehow to reduce these three C and then we're going to start talking
about much more advanced cache organizations and optimizations we can
make. So, let's briefly talk about it. It's sort of a hodgepodge.
And this is kind of just the nature of the beast.
There's lots of little things that sort of add up to make either the cache more
effective or integrate the cache into your pipeline more, more efficient.
And this is only the beginning. We will have about five or six more topics
the next time. But let's start off by looking at what
we're going to be talking about today. So, we're going to look at how to build,
how to pipeline a cache right. So, one of the things we haven't really
thought about when we were working on our labs is we do it right to the cache and it
just sort of magically gets stored there. Or, or, but if you're actually trying to
build the real cache, you have to look at the tag data and then you have to write
the data because you don't want to be writing something into the cache if it's,
let's say, a different piece of data is in that same location.
So, you need to check the tag first. And we'll talk about some ways to optimize
that. We'll talk about a write buffer.
A write buffer is a extra data structure we can put after our first level of cache,
which can hold victims. So, victim is a line or a cache block,
that is being evicted from the cache. And if you're going to, let's say,
Replace something in the cache, you need some place to put the old data and, or the
data that was, let's say, dirty in the cache from before.
And the slow way to do this with things, we start with that up to this point is you
take that line and you wait for it to get into either memory or the next level of
cache. But if you want to go faster, you can
think about trying to store in some place inside the structure and we'll say, write
to the next level of cache or write to the main memory in the background.
We'll talk about multi-level caches. So, multi-level caches, meaning you have a
level one, a level two, a level three, maybe a level four cache, then main
memory, and why this can improve performance.
We'll talk about victim caches. Victim caches this is the idea that can
basically increase the associativity of a cache by putting a low associative cache
or maybe a direct mapped cache next to a cache with very high associativity.
So, maybe like a fully associative cache and they use them together to basically
implement a, a pseudo higher associativity cache.
Cough We're going to talk today about prefetching, both hardware and software.
So, prefetching is the act of either a piece of hardware or software trying to
get data that you need early into your cache.
We'll talk about how to increase the bandwidth to your cache.
So, up to this point, we've looked at doing one memory operation per cycle.
Well, we built machines that can execute two operations per cycle, and on your
exam, you had a machine which could execute three operations per cycle.
Wouldn't it be great to be able to do three memory operations per cycle versus
just three ALU operations? So, how do we, how do we go about doing
that? Well, we can put multiple ports on this
structure, but that's not particularly great for performance from, from a clock
perspective, because it makes the structure very large.
So, you can start to think about banking which we'll talk about today as a, as a
mechanism to increase the bandwidth. And then, we'll talk a bunch, about a
bunch of different compiler optimizations or software optimizations that a compiler
can do to restructure code to have better cache performance.
So, this is just what we're talking about today.
Next lecture, we're going to continue on this.
And the main topic of next lecture as far as events, cache optimizations are
concerned, Is we're gonna be talking about how to
build caches which can deal with out of order memory or can deal with the ability
to have loads and stores which hit in the cache, or loads and stores which miss in
the cache, and then you can continue on past that point.
So, in our sequential machine up to this point, we've all, even in our out of order
machines we were talking about, when you took a cache miss, you basically had to
stop the machine. Cuz you have to wait for the data to come
back cuz you didn't know if the next instruction was going to need this data or
the next, maybe the next load will need that data or something like that.
So, if you want to try to go out of order, you can start thinking about that and
that's what the, the bulk of next of next lecture will be about, is how you go out
of order in your memory system or this what, this is largely sometimes called is
a nonblocking cache. Cough And I should point out that a
nonblocking cache is actually not just for out of order processors.
This is sometimes good for in order processors.
So, if you have a in order processor, the instruction sequence is going in order but
you don't want to, let's say, wait when you take a cache miss.
You could actually have the memory system be out of order with the pipeline be in
order. That's actually a pretty common thing for
people to do. Okay, so here, here let's go look at our
review. We have processor.
And in our basic machine, we go out to main memory when we want to get data.
Main memory is big. It holds everything, but it can be slow.
So, we put a cache in front of that, which keeps a subset of all the data in the main
memory. And it has a faster time and can take
something, for instance, like this example here,
Where you go to access memory, and you have to wait for the memory to come back.
And it can shrink this time. If we look at that as sort of, you know,
average memory access, It's your hit time plus your miss time
times your or assume your miss rate times your miss penalty.
We, this is just review. Okay.
So, going back to the three Cs, here we have a four-block cache. So, say, to a
side of associated cache and let's look at one of the major types of misses you can
So, this just means that's the first time any, any portion of your program has gone
to go access the data. There's no real way around this.
Maybe you can prefetch, but, you know, you're not really going to magically be
able to get the data in there early, unless you have some sort of aggressive
prefetching mechanism. Capacity.
So, in this cache here, we have four blocks.
If you're looping over, let's say, five cache box worth of data,
You're basically not going to ever be able to fit all the data in this cache or in
more common cases, let's say, you have a cache that's eight kilobytes cache, that's
your 01 cache and you're trying to add two arrays where each array is sixteen
kilobytes. Well, that means 32 kilobytes of data,
maybe even another eight kilobytes of data if you're not doing it in place for the
result matrix and it's just not going to fit in your cache.
You have capacity problems and things are just basically going to be continually
kicked out of your cache. Conflict means that you don't actually
have you, you have enough space in your cache, but roo many different loads or
stores you're trying to access alias to the same location in your cache.
So, let's say, you're trying to loop over three different cache blocks or trying to
access three different cache blocks which all alias to one set in the cache.
And effectively you can only fit two things and you're trying to fit three
things in two spots. And this is kind of a pigeon hole
principle here comes, comes up and says, no you just can't fit three things in two
boxes. So, we get conflicts.
Okay, so still, still review here. Let's, let's take a look at one way to
reduce our hit time. Hit time is just, the, the fast case.
That's actually you find the data in your cache.
This is the good case. But for hit time,
One way to make it lower, this is actually in nanoseconds, is just to have a smaller
cache or a simpler cache. So, small, simple caches are in this
darwing here. We have time on the vertical axis.
On the x-axis we have different cache sizes.
And the 16-kilobyte cache is faster than, let's say, the 256-kilobyte cache.
So, if you make a smaller cache, you can go faster.
And we've seen examples of this, actually, in the Pentium four when we were talking
about super pipelining. In the Pentium four, they had a very small
cache and the real reason they did that is because they had such a high clock
frequency that they couldn't put anything bigger and still fit it in one clock
cycle. You know, in a perfect world, if your
clock cycle was, was slower, you could try to fit some bigger cache in there.
So, you can try to reduce the miss rate. there's a couple different techniques on
how to reduce the miss rate and this is, this is the advance technique which we
will talk about later. But one of the basic ones is to change the
block size. As you make the box size larger, you'll be
pulling in more data so you'll get spatial locality and hopefully, you'll be pulling
in data that you actually need. So, the, you know, the, the positives of
this is you're really going to have less tag overhead if you have a larger block
size or cache line size. Because for the same amount of data, let's
say, you got a 64-byte line versus a 128-byte line, you're going to have half
as much tag sort of cost or, or area dedicated to the tags.
Cough And, and usually, as you go to larger block sizes,
This is good for the farther out layers of your cache, because you can basically
burst more data in. So if you look at something like a DDR2
memory or DDR3 memories. This is the traditional DRAM technologies
that we are using today. They want to actually, their block size,
their line size is very, very long, sort of thousands of bits. And if you have a
longer block size, you can use all those bits and pull them all into your cache in
one big chunk, and you can use wider buses.
Cough Downside? You can waste bandwidths.
If you have a very large block, and you don't actually use all the data in the
block, Let's say, you use one byte out of your,
out of 256-byte cache, You're just pulling all the data and
you're wasting lots of off chip bandwidths for no real good reason.
And if you have fewer blocks, You're going to have less locations to put
data for the fixed sized cache, so you might actually increase your conflict
rate. So, this, this chart here,
What we're trying to show this is from your book,
We're plotting the block size versus the miss rate.
And you want the lowest miss rate possible.
So, if you start here with a 16-byte block or 16-byte cache size or cache line.
As you increase it, the miss rate actually goes down.
But at some point, you end up wasting a lot of bandwidth cuz you don't have enough
of, of the, the program will ,, it does not have enough spatial locality.
So, you actually just end up pulling in more data than you need.
Cough You also could potentially have more conflicts.
So it starts to go up. So, there's sort of a minimum here.
Traditionally, people have thought about this as actually, increasing block size,
usually increases your performance for most applications.
But that's up to a limit. And the reason why people traditionally
thought this is, cache size is used to be sort of smaller or I mean, cache block
sizes or line sizes used to be smaller and they sort of got, starting getting bigger,
and then people really weren't looking way out here.
applications as you get farther out. Okay.
So, another way we, Just the basic ways we can reduce miss
rate, You could just have a bigger cache.
This one is great. Bigger cache, lower miss rate.
You got to pay area for this. And potentially it takes longer to access
this larger cache, as we'd seen in the first graph that we showed today.
I'm not going to really talk about this, but empirical rule of thumb.
If the cache size is doubled, your miss rate usually goes down by a square root of
two. It's just empirical, people have found,
but good rule of thumb to know. Same, same graph, tells a different story.
If you go to a higher associativity, your miss rate goes down.
So, if you have a direct mapped cache, and you go to a two-way set of cache,
And this rate goes down. And the traditional sort of rule of thumb
about this is if you have a direct mapped cache of size n.
It has roughly the same miss rate as a two-way set of associative cache that is
half the size. You might say, wow that's, that's pretty
amazing so if I just continually apply that over and over again, it doesn't quite
work. Laugh it also gets very hard to actually
build very highly associative caches cuz your, your, your tag check becomes pretty
hard. Because if you have to check against all
different places in the cache, it could be imparallel, and that becomes hard.
