Lecture 25: WrapUp

1
Lecture 25 Wrap-Up

• Mid-term-II stats
• High 91
• Mean 73.12
• Qs 1-3 half the class got 25/25
• Qs 4 only one student got 25/25 almost no one
  mentioned
• that well need a mechanism to determine
  exclusivity
• Qs 5 highest was 22/30 very few mentioned that
  allowing
• blocks to move would complicate search

2
Example Solutions
3
Example Solutions
4
Example Solutions
5
Example Solutions
6
Example Solutions
7
Example Solutions
8
CPU 2
CPU 3
L1D
L1I
L1D
L1I
CPU 4
L1I
L1D
L1D
CPU 1
L1I
CPU 5
L1I
L1D
L1D
CPU 0
L1I
CPU 6
L1D
L1I
CPU 7
L1D
L1I
9
Tetris?!
10
Non-Uniform Cache Access (NUCA)

• Many open problems in NUCA and D-NUCA
• How should search happen?
• Allocation/replacement/migration policies
• Managing bandwidth/latency on the network
• Prefetch mechanisms
• Selective replication of blocks
• Efficient write-throughs
• Power/performance trade-offs
• P.S. We have simulators, etc., to help model
  such
• caches in case anyone is interested

11
Shameless Plug

• CS 7810 Advanced Architecture
• Lectures based on seminal (and still relevant)
  papers
• Not much work, apart from class project (in
  teams)
• Class project can involve as little as 1 weeks
  worth of
• concentrated effort
• or, enough to get a paper out of it
• you WILL work on novel problems
• lots of help from me/other students with the
  simulator

12
3-D

• Imagine a similar problem in 3D

C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
13
3-D

• Imagine a similar problem in 3D

C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
Must schedule threads to manage temperature
14
Single Thread Performance

• To improve single-thread performance, can even
  schedule
• a single threads instructions across cores
  large window
• of in-flight instructions to mine high ILP
  requires high
• levels of speculation (power-hungry!) any
  solutions?

C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
C
P
15
Heterogeneous CMPs (Alpha EVx and Cell)
in-o
o-o-o
o-o-o
16
NASCAR Applied to CPUs !?!

• Bullet

Source Eric Rotenberg (NCSU)
17
Runahead Execution
Single thread in a baseline architecture
Single thread executing in tandem with a helper
thread
18
Reliability
For power
For performance
P1
C2
P2
C1
SMT core 1
SMT core 2
19
Title

• Bullet