ruby: set: reimplement using std::bitset

Information
Submitter:	Nilay Vaish
Repository:	gem5
Branch:	default
Bugs:
Depends On:
Reviewers
Groups:	Default
People:

Description

Changeset 11077:0dfa7e7d62d0
---------------------------
ruby: set: reimplement using std::bitset
The current Set data structure is slow and therefore is being reimplemented
using std::bitset. A maximum limit of 64 is being set on the number of
controllers of each type.  This means that for simulating a system with more
controllers of a given type, one would need to change the value of the variable
NUMBER_BITS_PER_SET

Testing Done

Issue Summary

Description	From	Last Updated	Status
m_bits[i].getSize() returns NUMBER_BITS_PER_SET, which is 64 for all machine types. In case of broadcast, all bits are set in m_bits. ...	Jieming Yin	Aug. 31, 2015, 7:59 a.m.	Resolved

I am glad to see this patch coming out because the current implementation is pretty confusing. However, would it be better if we keep the flexibility of dynamically adjusting the length of the bitset instead of hardcoding NUMBER_BITS_PER_SET? In some systems the number of computation units can go beyond 64, but with much fewer LLCs or dirs. That is my main concern.

Nilay Vaish Aug. 14, 2015, 6:13 p.m. (Aug. 14, 2015, 6:13 p.m.)

What you are asking for is not possible with std::bitset since it requires the size to be fixed at compile time. Another option could be to use std::vector<bool>. Somehow my feeling is that vector<bool> is going to slower than bitset. I would say just raise the number of bits per set to whatever you need. Even if you have a small number of shared controllers, I think bitset would be faster than the current implementation.

Ship It!

src/mem/ruby/common/NetDest.cc (Diff revision 1)

m_bits[i].getSize() returns NUMBER_BITS_PER_SET, which is 64 for all machine types. In case of broadcast, all bits are set in m_bits. Will it cause problem if the number of actual components is smaller than NUMBER_BITS_PER_SET? Perhapse in class Set, add a size parameter is a better idea to prevent such situation.

Show all issues

Nilay Vaish Aug. 31, 2015, 7:59 a.m. (Aug. 31, 2015, 7:59 a.m.)
```
I have added back the m_nSize variable.
```

Tony Gutierrez Sept. 3, 2015, 11:05 a.m. (Sept. 3, 2015, 11:05 a.m.)

Maybe I'm missing something here, but why is m_nSize necessary for broadcast? It seemed to me that broadcast() only looks to see if all bits for the set are set, and doesn't determine how many to components to broadcast to. So it seems fine to me to have the logic of broadcast() be such that it will be considered a broadcast if all bits are set, and not a broadcast otherwise, the fact that there are extra bits set doesn't seem to matter.

Nilay Vaish Sept. 3, 2015, 1:56 p.m. (Sept. 3, 2015, 1:56 p.m.)

On Thu, 3 Sep 2015, Tony Gutierrez wrote:

Maybe I'm missing something here, but why is m_nSize necessary for
broadcast? It seemed to me that broadcast() only looks to see if all

Because of the way broadcast is implemented by the on-chip network.

bits for the set are set, and doesn't determine how many to components
to broadcast to. So it seems fine to me to have the logic of broadcast()
be such that it will be considered a broadcast if all bits are set, and
not a broadcast otherwise, the fact that there are extra bits set
doesn't seem to matter.

The on-chip will try to send the broadcast message to controllers corresponding
to bits that are set. Hence, if there is no size variable to determine the maximum
number of bits in use, then messages will be sent to non-existent controllers.

Ship It!

Description:

~		Changeset 11053:07ad422ebbd9
	~	Changeset 11077:0dfa7e7d62d0

		ruby: set: reimplement using std::bitset
		The current Set data structure is slow and therefore is being reimplemented
~		using std::bitset. The maximum limit of 64 on the number of controllers of
~		each type is being retained. This means that for simulating a system with
~		more controllers of a given type, one would need to change the value of the
~		variable NUMBER_BITS_PER_SET
	~	using std::bitset. A maximum limit of 64 is being set on the number of
	~	controllers of each type. This means that for simulating a system with more
	~	controllers of a given type, one would need to change the value of the variable
	~	NUMBER_BITS_PER_SET

Diff:

Revision 2 (+141 -422)

Show changes

	src/mem/ruby/common/NetDest.hh
	src/mem/ruby/common/SConscript
	src/mem/ruby/common/Set.hh
	src/mem/ruby/common/Set.cc

Ship It!

You have a pending review.

Review Board 2.0.15

This change has been marked as submitted.

Screenshots

Files

Issue Summary

Description:

Changeset 11053:07ad422ebbd9

Changeset 11077:0dfa7e7d62d0

Diff:

Status: Closed (submitted)