1. Physical Design of Snoop-Based Cache Coherence in Multiprocessors
Muge Guher

2. Cache Coherence Definition
A Microprocessor is coherent if the results of any execution of a program can be reconstructed by a hypothetical serial order.
Write propagation
Writes are visible to other processes
Write serialization
All writes to the same location are seen in the same order by all processes (to “all” locations called write atomicity)
E.g., w1 followed by w2 seen by a read from P1, will be seen in the same order by all reads by other processors Pi

3. Cache Coherence Snooping Shared memory multiprocessor environment
Main Memory is passive
Caches distribute state transitions to other caches and memory
All caches listen to snoop messages and act on them
Most machines use cache coherence protocols with different trade-offs
But, performance (latency and bandwidth) also depends on physical implementation.
Bus design
Cache design
Integration with memory

4. Cache Coherence Requirements
Protocol Algorithm
States
State transitions
Actions/Outputs
Physical Design
Protocol intent is implemented in FSMs
Cache controller FSM
Multiple states per mis
Bus controller FSM
Other Controllers
Support for:
Multiple Bus transactions
Multi-Level Caches
Split-Transaction Busses

5. Design Wish List Implementation should be Correct
Require minimal extra hardware
Offer high performance
High Performance can be achieved with multiple events in progress, overlapping latencies
Leads to numerous complex interactions between events
More bugs!

6. Design Issues with implementing Snooping
Cache controller and tags
Bus side and processor side interactions
Reporting snoop results: how and when
Handling write-backs
Non-atomic state transitions
Overall set of actions for memory operation are not atomic
Race conditions
Atomic operations
Deadlock, livelock, starvation, serialization.

7. Cache Controller and Tags
Must “monitor” bus operations and “respond” to processor operations
Two controllers: bus-side, and processor-side
Bus transactions: Bus-side capture address and perform tag check.
Fail, snoop miss, no action
Hit, cache coherence protocol, RMW on state bits
For single level caches, duplicate set of tags and state or dual-ported tag and state store
Controller is an initiator and responder to bus transactions.
Data is not duplicated
Both sets of tags may be updated simultaneously
Single-level snoopy cache organization [1]

8. Reporting Snoop Results How does memory know another cache will respond and provide a copy of the block so it doesn’t have to?

Uniprocessor
Initiator places an address on the bus
Responder must acknowledge within a time-out window (wired-OR), otherwise bus error.
Snooping Caches
All caches must report on the bus before transaction can proceed.
Snoop result informs main memory, if it should respond or a cache has a modified copy of the block.
When and how the snoop result is reported on the bus?
For Example to implement MESI protocol,
Memory needs to know; Is block dirty? Should it respond or not?
Requesting cache needs to know; Is block shared?

9. When to report Snoop Results
Within fixed number of clock cycles from the address issue on the bus
Dual set of tags, high priority processor access to the tags.
Both set is inaccessible during processor updates.
Extra HW & longer snoop latency, but simple memory subsystem
Pentium Pro, HP Servers, Sun enterprise.
After a variable delay
Memory assumes one of the caches will supply the data, until all have snooped and indicated results.
Easier to implement, tag access-conflicts and high performance
Higher performance, don't have to assume worst case delay
SGI Challenge, fetches the data and stalls until snoop complete
Immediately
Main memory maintains a state bit per block, modified in a cache.
Complexity introduced to main memory subsystem

10. How to Report Snoop Results
Three wired-OR signals
1,2 : Two for snoop results
Shared: asserted if any cache has a copy
Dirty: asserted if some cache has a dirty copy
Dirty cache knows what action to take
3 : One indicating snoop valid. Inhibit signal, asserted until all processors have completed their snoop.
Illinois MESI protocol allows cache-to-cache transfers.
Retrieve data from other caches rather than memory.
Priority scheme needed
SI Challenge, Sun Enterprise Server, only in exclusive or modified state.
Challenge updates memory during cache-to-cache transfer (no shared modified state)

11. Single Level Snooping Cache
Assumptions
- Single Level write-back cache
- Invalidation protocol
- Processor can have one memory request outstanding
- The System bus is atomic
Snooping cache design [1]

12. Multi-level Cache Hierarchies How would a design of a cache controller be modified in case of L1/L2 caches?
Complicates Coherence
Changes made by the processor to L1 cache may not be visible to L2 cache controller, which is responsible for bus operations
Bus transactions are not directly visible to L1 cache
A Solution:
Independent bus snooping HW for each cache level hierarchy
L1 cache is usually on the processor, on chip snooper consumes pins to monitor shared bus
Duplicating tags consumes too much on chip area
Duplication of effort between L1 and L2 snoops.
Intel’s 8870 chipset has a “snoop filter” for quad-core

13. How do you guarantee coherence in a multi-level cache hierarchy?
Better Solution: Based on “Inclusion Property”
If memory block is in L1 cache it must also be in L2 cache
If the block is in modified state (or shared-modified) in L1 cache, then it must also be marked modified in L2 cache, (its copy in L2)
Therefore:
only a snooper at L2 is necessary, as it has all the required information
If a BusRd requests a block that is in modified state in either cache, then L2 can wave memory access and inform L1.
Now information flows both ways:
L1 accesses L2 for cache miss handling and block state changes;
L2 forwards to L1 blocks invalidated/updated by bus transactions;

14. Inclusion Property
Difficulties with maintaining the inclusion property:
L1 and L2 may have different eviction algorithms (replacement differences)
While a block is kept by L1 it may be evicted by L2
Separated data and instruction caches.
Different cache block sizes.
On a most commonly encountered case, inclusion works automatically:
L1 is direct mapped
L2 is either direct mapped or set associative
Same block size for both caches
Number of sets in L1 is smaller than in L2

15. Explicitly Maintaining Inclusion
Extend the mechanisms used for propagating coherence events to cache hierarchy.
Propagate L2 replacements to L1
Invalidate or flush messages
Propagate bus transactions from L2 to L1
Send all transactions to L1 (even if the given block is not present there)
Add extra state to L2 (a bit per block) which blocks in L2 are also in L1 (inclusion bit)
On write: propagate modified state from L1 to L2. If L1 is:
Write-through (so all modifications affect also L2), invalidate
Write-back :
Add per bit state every block in L2, "modified-but-stale"
Request flush from L1 on Bus read
L2 serves as a filter for the L1 cache, screening out irrelevant transactions from the bus, i.e. dual tags are less critical with multilevel caches

16. Propagating transactions for Coherence in Hierarchy
How is the transaction propagated for multilevel caches? Show some examples of modern systems.

Only one transaction on the bus at a time.
Transactions are propagated up and down the hierarchy, bus transactions may be held until propagation completes.
Performance penalty for holding processor write until BusRdX has been granted in high, so motivation to de-couple these operation
Two-level snoopy cache organization[1]

17. Split Transaction Bus
In a Split-transaction bus (STB), transactions that require a response are split in two independent sub-transactions: a request transaction and a response transaction.
Arbitrate each phase separately
Other transactions are allowed to intervene between request & response
Buffering between bus and the cache controllers allows multiple outstanding transactions (waiting for snoop and/or data responses)
Pro: By pipelining bus operations the bus is utilized more efficiently.
Con: Increased complexity.
Mem Access Delay
Address/CMD
Data
Bus
arbitration

18. Issues supporting STBs
A new request can appear on the bus before the snoop and/or servicing of an earlier request are complete;
these requests may be conflicting requests (same block);
The number of buffers for incoming requests and potential data responses from bus to cache controller is usually fixed and small, flow control is needed
Since requests from the bus are buffered, when and how snoop and data responses are produced on the bus
In the same order as requests arrive?
Snoop and data response together or separately
Example separately: Sun, together: SGI
There are 3 phases in a transaction:
A request is put on the bus
Snoop results are sent by other caches
Data is sent for the requesting cache, if needed

19. SGI Challenge Example Features:
Does not allow conflicting requests for same block (8 outstanding requests)
NACK Flow-control
NACK as soon as request appears on bus, requestor retries
Separate command (incl. NACK) + address and tag + data buses
Responses may be in different order than requests
Order of transactions determined by requests
Snoop results presented on bus with response
Examine implementation specifics of:
Bus design, request response matching
Snoop results
Flow Control

20. Two independently arbitrated buses:
How would a design of a cache controller be modified in case of split transactions buses? Show examples of modern split transaction buses/systems? How many outstanding transactions are allowed?

Bus Design
Two independently arbitrated buses:
Request: command+address (BusRd, BusWB + target address)
Response: data
Match each response to outstanding request, since they arrive out of order
Tag request with 3-bits (8 outstanding) when launched
Tag arrives back with corresponding response
Address bus is free, as tag is sufficient for request matching
Address and data buses can be arbitrated seperately
Separate bus lines for arbitration, flow control and snoop results

21. Bus and Cache Controller Design
To keep track of outstanding requests on the bus:
each cache controller maintains eight entry buffer, “request table”
A new request on the bus, added to all request tables at same index,
Index is 3-bit tag assigned at arbitration
Table entry contains; block address, request type, state in that cache etc.
Table is fully associative, new entry can be placed anywhere in table
Checked for a match by the requesting processor and by all snooped requests and responses on the bus
Entry and tag freed when response is observed on the bus,
Now tag can be reassigned by bus

22. Bus Interface and Request Table
Bus interface logic to accommodate split-transaction bus [1]

23. Snoop Results & Request Conflicts
Variable delay snooping
Snoop portion of the bus consists of three wired-OR lines
Sharing, dirty, inhibit
Request phase determines who will respond, but may take may cycles and intervening request response transactions
All controllers present their snoop results on bus when they see response
No data response or snoop results for write backs and upgrades
Avoid conflicts by:
Every controller keeps record of pending reads in request table
Don't issue request for a block with outstanding response
Writes performed during request phase
However does not ensure sequential consistency!

24. Flow Control Implement flow control at:
incoming request buffers from bus to cache controller (write-back buffer)
Cache subsystem has a response buffer (address + cache block of data)
limit number of outstanding requests
Flow control is also needed at main memory,
Each of the 8 pending requests can generate a write-back to memory
Can happen in quick succession on bus
SGI Challenge: separate NACK lines for address and data buses
Asserted before ack phase of request (response) cycle is done
Request (response) cancelled everywhere, and retries later
Backoff and priorities to reduce traffic and starvation
SUN Enterprise: destination initiates retry when it has a free buffer
source keeps watch for this retry
guaranteed space will still be there, so only two “tries” needed at most

25. Preventing violation of Sequential Consistency
SC: Serialization of operations to different locations.
Multiple outstanding requests on the bus, invalidations are buffered between bus and cache and are not applied to cache immediately
Commitment versus completion
Value produced by a write commit may not be visible to other processors
Condition necessary for SC: a processor should not be allowed to actually see the new value to a write before previous writes (in bus order) are visible to it.
not letting certain types of incoming transactions from bus to cache be reordered in the incoming queues
allowing these re-orderings in the queues, but then ensuring that the important orders are preserved at the necessary points in the machine.
a simpler approach is to threat all the requests in FIFO order. Although this approach is simpler, it can have performance problems;

26. Multi-level Caches and STB
Considerable number of cycles for a request to propagate through cache hierarchy
Allow other transactions to move up and down hierarchy while waiting
To maintain high bandwidth while allowing the individual units (controllers and caches) to operate at their own rates, queues are placed between levels of the hierarchy.
Leads to deadlock and serialization issues

27. Deadlock Fetch deadlock:
Must buffer incoming requests/responses while request outstanding
One outstanding request per processor => need space to hold p requests plus one reply (latter is essential)
If smaller (or if multiple o/s requests), may need to NACK
Then need priority mechanism in bus arbiter to ensure progress
Buffer deadlock:
L1 to L2 queue filled with read requests, waiting for response from L2
L2 to L1 queue filled with bus requests waiting for response from L1
Latter condition only when cache closer than lowest level is write back
Could provide enough buffering, requires a lot of area, not scalable
Queues may need to support bypassing

28. Sequential Consistency
Separation of commitment from completion even greater with multi level cache
Do not wait for an invalidation to reach all the way up to L1 and return a reply, consider write committed when placed on the bus
Fortunately techniques for single-level cache and ST bus extend, either method works:
don’t allow certain re-orderings of transactions at any level
don’t let outgoing operation proceed past level before incoming invalidations/updates at that level are applied

29. Shared Cache Designs Are there any solutions of shared L2 caches that are based on bus network? How does the bus network need to be modified to support shared caches?
Benefits of sharing a cache:
Eliminates the need for cache-coherence at this level
If L1 cache is shared then there are no multiple copies of a cache block and hence no coherence problem
Reduces the latency of communication.
L1 communication latency 2-10 clocks, main-memory many times larger
reduced latency enables finer-grained sharing of data
Pre-fetching data across processors.
With private caches each processor incurs miss penalty separately
Reduces the BW requirements at the next level of the hierarchy.
More effective use of long cache blocks, as there is no false sharing;
Shared cache is smaller than the combined size of the private caches if working sets from different processors overlap

30. Shared Cache Designs Extreme case:
All processors share a L1 cache, below is a shared memory
Processors are connected to shared cache by a switch, More likely a crossbar to allow cache access by processors in parallel
Support high BW by interleaving cache and main memory
Disadvantages of sharing L1:
higher bandwidth demand
hit latency to a shared cache is higher than to a private one
higher cache complexity
shared caches are usually slower
instead of constructive interference (like the working set example), destructive interference can occur

31. Example of Shared Cache Designs
Alliant FX-8 machine (1980's),
8 custom processors
Clock cycle 170ns
Processors connected using crossbar to 512Kbyte, 4-way interleaved cache
Cache: 32 byte block size, direct mapped, write-back, 2 outstanding misses per processor
Encore Multimax (contemporary)
Snoopy cache coherent multiprocessor
Each private cache supports two processors instead of one
Practical approach:
private L1 caches and a shared L2 cache among groups of processors.
packaging considerations are also important

32. References
[1] David Culler, Jaswinder Pal Singh, and Anoop Gupta, Morgan Kaufmann, Parallel Computer Architecture: A Hardware/Software Approach, Morgan Kaufmann; preliminary draft edition (August 1997), pp
[2] Daniel Braga de Faria, Stanfard, Book Summaries, retrieved October 2010, from students.stanford.edu/~dbfaria/
[3] Andy Pimentel, “Introduction to Parallel Architecture”, retrived on October 2010, from pp
[4] R. H. Katz, S. J. Eggers, DA.A. Wood, C.L Perkins and R.G. Shedon, “Implementing a cache Consistency Protocol”, Proceedings of the 12th ISCA, 1985, pp
[5] M. S. Papamarcos, J.H. Patel, “A low Overhead Coherence Solution for Microprocessors with Private Cache Memories” , Proceedings of the 11th ISCA, 1984, pp
[6] R. Kumar, V. Zyuban, and D. M. Tullsen, “Interconnections in Multi-Core Architectures: Understanding Mechanisms, Overheads and Scaling”. In ISCA, Jun 2005.