In this course, you will learn to design the computer architecture of complex modern microprocessors.
More Cache Coherence Protocols
Loading...
From the course by Princeton University
Computer Architecture
283 ratings
From the lesson
Multiprocessor Interconnect 1
This lecture covers the design of interconnects for a multiprocessor.
Meet the Instructors
David WentzlaffAssociate Professor
Electrical Engineering
So, today we're going to continue our adventure in computer architecture and
talk more about parallel computer architecture.
last time we talked about coherence, memory coherence, and cache coherence,
systems and to differentiate that from memory consistency models which is a
model of how memory is supposed to work, versus the underlying algorithms that try
to keep memory consistent, and try to implement the consistency models.
We left off last time, we, we were talking about MOESI, or also known as the
Illinois protocol, and we walked through all of the
different arcs through here. And if you recall what we were talking
about, was, we split the shared state from the MSI protocol into two states,
shared and exclusive. And the insight here is, it's very common
for programs to read a memory address, which will pull it into your cache.
And then go modify that memory address. So for instance, if you want to increment
a number. You're going to do a load.
It's going to bring it into your, your, ca-, or into your register set.
But also into your cache. You're going to going to increment the
number and then you do a write back to the exact same location.
Pretty common in imperative programming languages.
Declarative programming languages like Scheme and such, they may at times copy
everything. But for declara- excuse me, imperative
programming languages it's pretty common to actually change state in place.
So, because of that, you can bring it right into, this exclusive stage, and
then when you have to go to modify it, you would have to go and broadcast in the
bus. You know, you would have to talk to
anybody and would loose, hm, effectively this intent to write message.
Then you would have to send otherwise across the bus and waiting for that
address to be snooped on the bus, or be seen by all the other entities on the
bus. note I, I say entities in the bus.
We've been talking primarily about, processors, the last day,
but there can be other entities on the bus that want to snoop the bus.
So examples sometimes include coherent, IO devices.
So, this isn't very popular right now,
but I think this will become much more popular as soon as we start to have, GPUs
or Graphics Processing Units or general purpose GPUs,
which will be sitting, effectively, very close to our processor on the same bus,
and will want to take part in the coherence traffic of the processor.
So it's going to want to basically read and write to the same memory addresses
that the processor is reading and writing,
and take part in the cash coherence protocol.
[COUGH] At a minimum, usually your IO devices need to effectively tell the
processor when its doing a memory transaction that the processor should
know about. So typically when you are moving data
from a IO device to main memory, that's going to have to effectively go
across the person. Everyone is going to have to validate
their cache's, you have to snoop the traffic, or they will all have to snoop
that memory traffic from the IO device. So we had talked about MOESI as an
enhancement to MSI. Well, we left off last time, and we were
going to talk about two more enhancements that are pretty common.
one is been used widely in AMD Opterons. I think they still use this in AMD.
I think they use something similar to this still in AMD, is my understanding.
and the idea is you add an extra state here,
which is called ownership, or the owned state.
And effectively, what this is, is it looks just like our MOESI protocol from
before. But now, instead of having data in the
modified stage, when you, let's say another processor needs to go access that
data, instead of having to send all that data back to main memory, and validate
that line out to main memory, and go fetch it back from main memory.
Instead, you can do direct cache to cache transfer.
This is of a, basically an optimization here.
So, you don't have to right back to data to main memory,
and in fact you can allow main memory to be stale.
And you can just transfer the data across the bus from the one cache to the cache
which needs it. So in this example here, we're going to
look at this edge here. So another processor wants to read the
data. So we see an intent to write to a
particular cache line and our processor currently has it in the modified state.
We see this other processors intent to write, and.
[COUGH] or excuse me. Intent to read.
And we're actually going to provide the data out of our cache,
and not write it back to main memory, and transition the line in our cache to
this owned state. The other processors can now take it in,
and take it in a shared state. So they will have it a read, read only
copy. Now, note this is only for, for
read-only, we'll talk about if another processor wants to write to the state in
a second. So we have it in its own state, and what
we're trying to do here is this processor is tracking that, that data needs to be
written back to main memory at some point.
That's the whole purpose of this state here,
is we've basically designated a processor which owns the data and owns the modified
state. So the processors which take at read only get it into the shared state,
and if they need to invalidate the line, they don't need to contact anybody.
Because they are having a share state, they have a read-only copy.
They don't need to make any bus transactions.
[COUGH] So if you think about it, if you actually want to effectively have one
core, read the state from other, read this dirty state from the other core,
and then in some points it goes in and just invalidates it in the, in the second
core. If the data is not up to date as in, it
would be in main memory, you lose the changes.
So, by one processor keeping it in the own state here, it keeps track that at
some point, if it never gets invalidated out of that processor's cache, it needs
to write that out to main memory, to keep it up, up to date.
Now, there's a couple other arcs here. you can transition from the own state
back to the modified state if the processor, which has it in the owned
state wants to go to a write. [COUGH].
It can't do that while it's in the owned state,
because while it's in the owned state, other processors may have shared copies
of it. So, when it needs to do that,
if it wants to do, P1 wants to do a write here, it needs to re-invalidate everyone
else's copies across the bus. So it's going to have to send an intent
to write for that line, and everyone else will snoop that
traffic, and transition to the invalid state.
And then, this processor will be able to transition to the modified state,
and now it's able to actually modify the data.
Okay. So we've got this arc here, which we sort
of already talked about, is that if you're in the owned state, anyone else
can get read only shared copies of it. [COUGH].
They can't go get an exclusive copy, because that would basically violate this
notion, because then they would be able to upgrade to modified without telling
anybody, and we don't want that. But they can get shared read-only copies
of the data and then there's this arc here from owned to invalid, is if some
other processor wants to write the data. We're going, processor 1, P1 here will
say, we'll see the intent to write from
another processor. It will, snoop that traffic effectively,
and at that point it will transition to this invalid state.
note here that this intent to write, we may need to provide information across
the bus when we're in the owned state. Because if the only, if we're the only
owner of that, or the only cache that has that data, and the other processor is
basically going straight into this state here via rightness, we're going to need
to provide the data. Okay, so, questions about MOESI?
So far, But a basic, extra optimization.'Cause we
don't have to. We can basically transfer data around.
And one cache can have a, a, a cache line in the owned state.
And later, some other cache, you know, the exact same cache line in the own
state. And it can basically bounce around
without ever having to go out to main memory.
And this, this decreases our bandwidth out to the main memory system.
Okay. Then we're going to talk about MESIF or
MESIF, which is actually used in the core I7 in the most up to date Intel
processors. And it looks very similar to MOESI,
except we're going to see an extra little letter in this one bubble here.
Effectively, the, what's going on here is we add an
extra state called the Forward State. And this is similar to sort of the
optimization we saw in MOESI, except it can't keep the data writeable.
So, what happens in this protocol is, let's say the first cache, which does a
read miss on a line for widely shared data is going to be elected and going to
get the data in this forward state. And then if other caches want to get read
only copies, bring it in shared. Instead of having to go out to main
memory, the cache that has it in the forward
state is going to provide that data across the bus.
So this is going to effectively decrease our bandwidth to main memory, by
providing the data out of another cache's,
cache, effectively, or another processor's cache rather, and then you
won't have to, have to transition it. Now, this is a little bit of a
simplification. There is a question here of, if you're in
this forward state and you invalidate the data.
Who has it? does anyone provide the data?
So there's sort of two choices here. One choice is no one has it in the
forward state. So when it's a snooper quest for a line,
it actually has, it just have to go out of the main memory.
That's kind of the easy case. The other case is you could try to
actually build a protocol where another cache when one, one cache invalids the
forward, it just chooses another cache. But probably the simplest thing you do is
when the forward, the forwarding core invalidates the data.
For whatever reason, you just go back out to main memory, because there's always a
copy in main memory. So effectively you're just keeping read
only copies. Yeah,
you're right. You're probably going to enter, into the
exclusive state. That's a good question.
so I read two different versions of this in, in different books.
So, I'm not quite sure Intel actually
documents what they do for, for this. probably what's okay, so, so, you
probably will, youre probably right. You probably want to enter straight into
exclusive state. If you have a read only copy,
What. You, you, yeah.
So what's going to happen is when you transition from E to S here, you're
going to transition from E to F. And then you're going to be able to make
this, you'll end up in the F state. So the first person who actually
downgrades is going to always end up in the F state.
but like I said. I saw other references where people said,
There were other people implementing something similar to this.
Where they, Some have, some have some election where
they figure out who is the, the forwarding,
Node, but probably the easiest thing to do is to downgrade from E to F.
So the rest of the course, we're going to look at how to scale beyond these
broadcast and these invalidate protocols that have to snoop on a bus.
so, some of the problems of building these.
Snooping systems is, that you need, it really affects how you design your
processor. So first of all, you're going to have to
add more bandwidth into your cache. Or at least more bandwidth into your tag
array. so one choice is going to dual port your
tags. Another choice is you can steal cycles
for snoops. So what I mean by steal cycles is if
there is a bus transaction happening and you need to check this against your tags,
you actually block the main processor that is associated with that cash from
accessing the cash that cycle, so you, you generate a stall signal to the cash
or to the main pipe. [COUGH] And one of the things here that
get a little tricky is, and this will affects your design, is if you have a
multilevel cache, usually you want to put your sort L2 tag array on the bus and
snoop against your L2 tag array. But if it hits there and you figure out
that you have to invalidate something. You're going to have to invalidate down
the entire cache hierarchy, all the way down to the level one cache.
So this can actually affect your throughput on your level one cache
effectively. And also, it sort of is, is annoying to
do, because it's going to effectively have to
reach down and touch your tag array of your L1 cache.
And as I had mentioned, I think, last time, briefly.
If you're thinking about something like a exclusive cache.
So a cache where the tags and the L2 don't have the tags in the L1.
You're going to have to check both tags for every snoop transaction, and that can
be pretty, pretty painful, to do. [COUGH] Or you have to copy, the L1 tags,
but it's effectively the same thing as just having a, inclusive cache,
but maybe for little less data storage. Okay, so what limits our performance?
Why can't we just build 1,000 processors on a big bus?
Well it's the same idea if you have 1,000 people in this room, and they're all
trying to shout to each other at the same time.
At some point you, you run out of both bandwidth, and more importantly you need
some way to coordinate them. But also, but also if you wanted to, if
you're required to basically serialize, the occupancy on the bus goes up.
So, if you have one bus with two people talking on the bus at a time, they each
can, let's say, and they talk 10% of the time, then you have a 20% utilized bus.
Well all of a sudden, if you have ten people on this bus, you have 100%
utilized bus and if you have 1,000 people, you have an oversubscribed bus
so, you have to worry about the bandwidth, and occupancy, because we do
need to make these different bus transactions atomic.
So it's not quite just a bandwidth problem.
And what, what I mean by balance, you could make the bus wider.
To increase the bandwidth, but it's not going to solve our problems.
Because there's an occupancy challenge here also that you need effectively
atomic transactions to happen across the bus in order to keep the cache coherence
protocol correct. Okay, so before we move off this topic
into our interconnection networks, that we were talking about today, hm, I want
to talk about one of the challenge of, that happens in simple cache coherence
systems. And that's false sharing.
So caches, like to track information on, a particular bloc size.
So, we've talked about caches which have 64 byte, lines, or 64 byte block sizes,
and they can be bigger or smaller than that.
Now, one of the things that happens that is pretty unpleasant in these coherence
protocols is let's say, you take a piece of data which is shared, and needs to be
coherent between two different processors.
And it's gets communicated relatively often.
And you put some other piece of critical data right next to it, on the same cache
line. All of the sudden, what's going to happen
is, because they're packed into one cache line, and we only track that information
on a per cache line basis, whenever that one piece of data, let's
say it's a four byte integer, and there's another four byte integer which is not
shared, [COUGH]. Whatever the first four by energy which
let's just say a lock or something like that gets, gets bounced around between
caches you're going to bounce around the other data.
So this can effectively can hurt your performance common case performance for
non shared data by having this true sharing of data happening.
And this is not something that typically happens in a normal cache because in a
uniprocessor cache system you're going to bring the data in and it's going to bring
everything in and you get spacial locality.
And if you. Pump it out, you know, you can, you can
get conflicts which are sort of equivalent to this but it's a little bit
different idea here. It's never going to be in the same line.
But with false sharing, we, we do see this.
Hm, now false sharing is interesting because people have come up with a whole
measure of techniques to avoid it. So, anyone have an idea, one, one really
good technique to avoid false sharing? What we can do, and this is pretty
common, is either the programmer or the compiler can detect that this is
happening and it will actually pad the information out.
So waste memory for highly contended pieces of data, and co-locate it with
nothing that is shared. [COUGH].
So one of the better examples of why you really have to care about this is
something like your stack. Sometimes if you, if you were to have,
let's say, a lock on your stack, there's a lot of data which you need to use
often, and it's all local. Stacks between threads are all local.
But if you have, like, some sort of variable that you pass to someone else,
which is a struct, and inside that struct is a lock, or something like that.
All of sudden, you're basically going to be bouncing a line around which is your
stack. And it's, other people are going to be
invalidating your stack. So one way to solve this is
when you put a lock, and the compiler can sometimes recognize
this. Because you can actually designate memory
addresses as locks, with special keywords, sometimes,
depending on the language. And when you do that, it'll say, oh,
don't put this with anything else, or maybe only collocate this data with
things that are other locks. because that may have bad sharing
performance anyway, for instance. So and really what you want to do here,
is not have a false sharing case. Now, the analog default sharing is
actually true sharing. So there are, there are cases where
you'll have multiple pieces of data that are, shared differently between different
lines. But they are also widely shared.
So example of this is, you have an array of locks, and different processors won't
be grabbing these blocks randomly. [COUGH] You can use similar techniques in
fold sharing. Now, you probably don't want all those locks to be on those cache
line. Because the locks are basically going to
be bouncing around, and everyone is going to be contending
for that one cache line to get it, modified in their ache,
or in the M state in their cache. So what you can think about doing is, is
actually just doing the similar technique, and putting, each of those
locks on a separate cache line. Okay, so let's switch gears here.
Explore our Catalog
Join for free and get personalized recommendations, updates and offers.
Coursera provides universal access to the world’s best education,
partnering with top universities and organizations to offer courses online.
© 2018 Coursera Inc. All rights reserved.
