1. Core 2 Duo die
“Just a few years ago, the idea of putting multiple processors
on a chip was farfetched. Now it is accepted and
commonplace, and virtually every new high performance
processor is a chip multiprocessor of some sort…”
Center for Electronic System Design
Univ. of California Berkeley
Chip Multiprocessors??
“Mowry is working on the development
of single-chip multiprocessors: one large
chip capable of performing multiple
operations at once, using similar
techniques to maximize performance”
-- Technology Review, 1999
Sony's Playstation 3, 2006
2. CMP Caches: Design Space
• Architecture
– Placement of Cache/Processors
– Interconnects/Routing
• Cache Organization & Management
– Private/Shared/Hybrid
– Fully Hardware/OS Interface
“L2 is the last line of defense before hitting the
memory wall, and is the focus of our talk”
3. Private L2 Cache
I$ D$ I$ D$
L2 $ L2 $ L2 $ L2 $ L2 $ L2 $
I N T E R C O N N E C T
Coherence Protocol
Offchip Memory
+ Less interconnect traffic
+ Insulates L2 units
+ Hit latency
– Duplication
– Load imbalance
– Complexity of coherence
– Higher miss rate
L1 L1
Proc
4. Shared-Interleaved L2 Cache
– Interconnect traffic
– Interference between cores
– Hit latency is higher
+ No duplication
+ Balance the load
+ Lower miss rate
+ Simplicity of coherence
I$ D$ I$ D$
I N T E R C O N N E C T
Coherence ProtocolL1
L2
6. Take Home Messages
• Leverage on-chip access time
• Better sharing of cache resources
• Isolating performance of processors
• Place data on the chip close to where it is used
• Minimize inter-processor misses (in shared cache)
• Fairness towards processors
7. On to some solutions…
Jichuan Chang and Gurindar S. Sohi
Cooperative Caching for Chip Multiprocessors
International Symposium on Computer Architecture, 2006.
Nikos Hardavellas, Michael Ferdman, Babak Falsafi, and Anastasia Ailamaki
Reactive NUCA: Near-Optimal Block Placement and Replication in Distributed Caches
International Symposium on Computer Architecture, 2009.
Shekhar Srikantaiah, Mahmut Kandemir, and Mary Jane Irwin
Adaptive Set-Pinning: Managing Shared Caches in Chip Multiprocessors
Architectural Support for Programming Languages and Operating, Systems 2008.
each handles this problem in a different way
8. Co-operative Caching
(Chang & Sohi)
• Private L2 caches
• Attract data locally to reduce remote on chip access.
Lowers average on-chip misses.
• Co-operation among the private caches for efficient
use of resources on the chip.
• Controlling the extent of co-operation to suit the
dynamic workload behavior
9. CC Techniques
• Cache to cache transfer of clean data
– In case of miss transfer “clean” blocks from another L2 cache.
– This is useful in the case of “read only” data (instructions) .
• Replication aware data replacement
– Singlet/Replicate.
– Evict singlet only when no replicates exist.
– Singlets can be “spilled” to other cache banks.
• Global replacement of inactive data
– Global management needed for managing “spilling”.
– N-Chance Forwarding.
– Set recirculation count to N when spilled.
– Decrease N by 1 when spilled again, unless N becomes 0.
10. Set “Pinning” -- Setup
P1
P2
P3
P4
Set 0
Set 1
:
:
Set (S-1)
L1
cache
Processors Shared
L2 cache
I
n
t
e
r
c
o
n
n
e
c
t
Main
Memory
11. Set “Pinning” -- Problem
P1
P2
P3
P4
Set 0
Set 1
:
:
Set (S-1)
Main
Memory
12. Set “Pinning”
-- Types of Cache Misses
• Compulsory
(aka Cold)
• Capacity
• Conflict
• Coherence
• Compulsory
• Inter-processor
• Intra-processor
versus
14. R-NUCA: Use Class-Based Strategies
Solve for the common case!
Most current (and future) programs have the following types of accesses
1. Instruction Access – Shared, but Read-Only
2. Private Data Access – Read-Write, but not Shared
3. Shared Data Access – Read-Write (or) Read-Only, but Shared.
15. R-NUCA: Can do this online!
• We have information from the OS and TLB
• For each memory block, classify it as
– Instruction
– Private Data
– Shared Data
• Handle them differently
– Replicate instructions
– Keep private data locally
– Keep shared data globally
16. R-NUCA: Reactive Clustering
• Assign clusters based on level of sharing
– Private Data given level-1 clusters (local cache)
– Shared Data given level-16 clusters (16 neighboring machines), etc.
Clusters ≈ Overlapping Sets in Set-Associative Mapping
• Within a cluster, “Rotational Interleaving”
– Load-Balancing to minimize contention on bus and controller
18. Just Kidding…
• Optimize for Power Consumption
• Assess trade-offs between more caches and more cores
• Minimize usage of OS, but still retain flexibility
• Application adaptation to allocated cache quotas
• Adding hardware directed thread level speculation
