CASHMERe

1
CASHMERe

• David Kunzman, Mei Chao, Esteban Pauli

2
Overview

• It is well accepted today that commercial
  workstations offer the best price/performance
  ratio and that shared memory provides the most
  desirable programming paradigm for parallel
  computing.
• CASHMERe provides a shared memory view of
  distributed memory machines.
• Provide two DSM coherence protocols
• Over message passing
• With DEC Memory Channel NIC

3
Data Sharing Management

• Page-sized coherence blocks for shared-memory
• Accesses detected by VM subsystem
• Each page
• A single distinguished home node master copy
• An entry in a global page directory maintains
  info about sharing state and home node location
• States Invalid, Read or Read-Write

4
Handling Page Faults

• Invalid page copy from home node via page update
  request
• Read access upgrade sharing state for the node
  and set permission to Read
• Write access
• Exclusive mode if no other shares
• Make a pristine copy (Twin), page ID into dirty
  list (Enable recovery later) Update sharing info
  and set permission to Read-Write

5
During Release/Acquire Operation

• Release
• Diff traverse processors dirty list and compare
  the working copy of each modified page to its
  twin
• After diff, write notices are sent to all sharers
  of the page
• Acquire
• Invalidate all pages with write notice

6
Programming in CASHMERe

• SPMD
• Primitives
• Locks acquire, release
• Barrier
• Flags
• acquire, release
• More advanced functionality
• void csm_wait_eq_flag(int index, long value)
• long csm_poll_flag(int index)
• http//www.cs.rochester.edu/research/cashmere/docs
  /csm_api.html

7
Example Initialization

1. struct int values4096 pGlobalData1
2. struct int others256 pGlobalData2
3. void main(int arg, char argv)
5. csm_init_memory_size(
6. csm_memory_page_round(sizeof(pGlobalData1
   ))
7. csm_memory_page_round(sizeof(pGlobalData2
   ))
8. )
9. csm_init_start(argc, argv) // spawn other
   threads
10. if (csm_pid 0)
11. pGlobalData1 csm_malloc(sizeof(pGlobalD
    ata1))
12. csm_distribute(pGlobalData1,
    sizeof(pGlobalData1))
13. pGlobalData2 csm_malloc(sizeof(pGlobalD
    ata2))
14. csm_distribute(pGlobalData2,
    sizeof(pGlobalData2))
16. csm_init_complete() // before this, only pe
    0 can touch global data
17. // compute using global variable...
18. pGlobalData1-gtvalues0 pGlobalData2-gtother
    scsm_pid
19. csm_exit(0)

8
Example 1D Jacobi

1. while(not converged)
2. my_start csm_pid chunk_size
3. my_end (csm_pid 1) chunk_size
4. for(i my_start i lt my_end i)
5. oldi curi // copy
6. for(i my_start i lt my_end i)
7. curi // compute
8. csm_barrier(0 / barrier ID /)

9
Example Dining Philosophers

1. while(still_hungry_for_pizza())
2. csm_lock_acquire(csm_pid)
3. csm_lock_acquire((csm_pid 1) csm_num_pid)
4. eat_pizza() // yummy
5. csm_lock_release(csm_pid)
6. csm_lock_release((csm_pid 1) csm_num_pid)
8. // can deadlock, just showing use of locks

10
Example flag use

1. // wait my turn
2. csm_wait_neq_flag(0, csm_pid)
3. csm_acquire_flag(0)
4. // take my turn
5. //let next person take their turn
6. cms_inc_flag(0, 1)

11
Example Bubble Sort

1. // Assume flag 0 is initialized to 0
2. for (int i 0 i lt N i)
3. csm_wait_neq_flag(0, csm_pid 2)
4. // swap with neighbor if needed
5. if (csm_pid 0 csm_pid 1)
6. csm_aquire_flag(0)
7. if (csm_pid 0) csm_inc_flag(0, 1)
8. if (csm_pid 1) csm_reset_flag(0)

12
Current Status

• Could not find a place to download it
• Last paper in 1999