SMP, Multicore, Memory Ordering & Locking
Transcript of SMP, Multicore, Memory Ordering & Locking
![Page 1: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/1.jpg)
SMP, Multicore, Memory Ordering & Locking
![Page 2: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/2.jpg)
These slides are made distributed under the Creative Commons Attribution 3.0 License, unless otherwise noted on individual slides.
You are free:to Share — to copy, distribute and transmit the workto Remix — to adapt the work
Under the following conditions:Attribution — You must attribute the work (but not in any way that suggests that the author endorses you or your use of the work) as follows:
“Courtesy of Kevin Elphinstone, UNSW”
The complete license text can be found at http://creativecommons.org/licenses/by/3.0/legalcode
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 3: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/3.jpg)
CPU performance increases are slowing
Computer Architecture A Quantitative Approach Fifth Edition John L. Hennessy, David A. Patterson
![Page 4: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/4.jpg)
Multiprocessor System• A single CPU can only go so fast
– Use more than one CPU to improve performance– Assumes
o Workload can be parallelisedo Workload is not I/O-bound or memory-bound
![Page 5: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/5.jpg)
Amdahl’s Law• Given:
– Parallelisable fraction P– Number of processor N– Speed up S
• 푆 푁 =
• 푆 ∞ =
• Parallel computing takeaway:– Useful for small numbers of CPUs (N)– Or, high values of P
o Aim for high P values by design
0
5
10
15
20
25
0 5 10 15 20 25 30
Speedup vs. CPUs
0.5 0.9 0.99
![Page 6: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/6.jpg)
• Classic symmetric multiprocessor (SMP)– Uniform Memory Access
o Access to all memory occurs at the same speed for all processors.
– Processors with local cacheso Separate cache hierarchy Cache coherency issues
Types of Multiprocessors (MPs)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
Cache
CPUCache
CPU Main Memory
Bus
![Page 7: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/7.jpg)
Cache Coherency• What happens if one CPU writes to address 0x1234 (and
it is stored in its cache) and another CPU reads from the same address (and gets what is in its cache)?– Can be thought of as managing replication and migration of data
between CPUs
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
Cache
CPUCache
CPU Main Memory
Bus
![Page 8: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/8.jpg)
Memory Model• A read produces the result of the last write to a
particular memory location?– Approaches that avoid the issue in software also
avoid exploiting replication for cooperative parallelismo E.g., no mutable shared data.
– For classic SMP a hardware solution is usedo Write-through cacheso Each CPU snoops bus activity to invalidate stale lineso Reduces cache effectiveness – all writes go out to the bus.
Longer write latency Reduced bandwidth
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
Cache
CPUCache
CPU Main Memory
![Page 9: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/9.jpg)
• NUMA MP– Non-uniform memory access
o Access to some parts of memory is faster for some processors than other parts of memory
– Provides high-local bandwidth and reduces bus contentiono Assuming locality of access
Types of Multiprocessors (MPs)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
Cache
CPU
Cache
CPU
Main Memory
Main Memory
Interconnect
![Page 10: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/10.jpg)
![Page 11: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/11.jpg)
Cache Coherence• Snooping caches assume
– write-through caches– cheap “broadcast” to all CPUs
• Many alternative cache coherency models– They improve performance by tackling above assumptions– We’ll examine MESI (four state)– ‘Memory bus’ becomes message passing system between caches
![Page 12: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/12.jpg)
Example Coherence Protocol MESIEach cache line is in one of four states
• Modified (M) – The line is valid in the cache and in only this cache.– The line is modified with respect to system memory—that is, the modified data in
the line has not been written back to memory.• Exclusive (E)
– The addressed line is in this cache only. – The data in this line is consistent with system memory.
• Shared (S)– The addressed line is valid in the cache and in at least one other cache. – A shared line is always consistent with system memory. That is, the shared state
is shared-unmodified; there is no shared-modified state.• Invalid (I)
– This state indicates that the addressed line is not resident in the cache and/or any data contained is considered not useful.
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 13: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/13.jpg)
Example
Cache
CPU
Cache
CPU
Main Memory
![Page 14: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/14.jpg)
MESI (with snooping/broadcast)
• EventsRH = Read HitRMS = Read miss, sharedRME = Read miss, exclusiveWH = Write hitWM = Write missSHR = Snoop hit on readSHI = Snoop hit on invalidateLRU = LRU replacement
• Bus TransactionsPush = Write cache line back to
memoryInvalidate = Broadcast invalidateRead = Read cache line from
memory• Performance improvement via
write-back caching– Less bus traffic
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 15: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/15.jpg)
• Each memory block has a home node
• Home node keeps directory of caches that have a copy– E.g., a bitmap of processors
per cache line
• Pro– Invalidation/update
messages can be directed explicitlyo No longer rely on
broadcast/snooping
• Con– Requires more storage to
keep directoryo E.g. each 256 bits of
memory (cache line) requires 32 bits (processor mask) of directory
Directory-based coherence
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
Computer Architecture A Quantitative Approach Fifth Edition John L. Hennessy, David A. Patterson
![Page 16: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/16.jpg)
Example
Cache
CPUCache
CPU
Main Memory
Cache
CPUCache
CPU
![Page 17: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/17.jpg)
• Chip Multiprocessor (CMP)– per-core L1 caches– shared lower on-chip caches– usually called “multicore”– “reduced” cache coherency
issueso Between L1’s, L2 shared.
Chip Multiprocessor (CMP)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
Bus
L1
CPU core
L1
CPU core Main
Memory
L2
CPU Package
![Page 18: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/18.jpg)
18
Cache lines can migrate between L1 caches belonging to different cores without involving the L2
Clean lines – DDI (Direct Data Intervention )
Dirty Lines – ML (Migratory Lines)
ARM MPCore: Cache-to-Cache Transfers
![Page 19: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/19.jpg)
19
Cache to Cache Latency
Significant benefits achievable if the working set of the application partitioned between the cores can be contained within the sum of their caches
Helpful for streaming data between cores may be used in
conjunction with interrupts between cores
Though dirty lines have higher latency they still have 50% performance benefit
![Page 20: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/20.jpg)
• D.M. Tullsen, S.J. Eggers, and H.M. Levy, "Simultaneous Multithreading: Maximizing On-Chip Parallelism," In 22nd Annual International Symposium on Computer Architecture, June, 1995
• replicated functional units, register state• interleaved execution of several threads
– As opposed to extracting limited parallelism from instruction stream.
• fully shared cache hierarchy• no cache coherency issues• (called hyperthreading on x86)
Simultaneous multithreading (SMT)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
L1 Cache
CPU Core
HW thread HW thread
![Page 21: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/21.jpg)
Summary• Hardware-based cache coherency:
– provide a consistent view of memory across the machine.– Read will get the result of the last write to the memory hierarchy
![Page 22: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/22.jpg)
Memory Ordering
• Example: a tail of a critical section/* assuming lock already held */
/* counter++ */
load r1, counter
add r1, r1, 1
store r1, counter
/* unlock(mutex) */
store zero, mutex
• Relies on all CPUs seeing update of counter before update of mutex
• Depends on assumptions about ordering of stores to memory
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
Cache
CPUCache
CPU Main Memory
Bus
![Page 23: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/23.jpg)
Memory Models: Strong Ordering• Sequential consistency
o the result of any execution is the same as if the operations of all the processors were executed in some sequential order, and the operations of each individual processor appear in this sequence in the order specified by its program
• Traditionally used by many architectures• Assume X = Y = 0 initially
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
CPU 0store 1, Xload r2, Y
CPU 1store 1, Yload r2, X
![Page 24: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/24.jpg)
Example
Cache
CPU 0
Cache
CPU 1
Main Memory
![Page 25: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/25.jpg)
Potential interleavings• At least one CPU must load the other's new
value– Forbidden result: X=0,Y=0
store 1, Xload r2, Ystore 1, Yload r2, XX=1,Y=0
store 1, Xstore 1, Yload r2, Yload r2, XX=1,Y=1
store 1, Yload r2, Xstore 1, Xload r2, YX=0,Y=1
store 1, Ystore 1, Xload r2, Xload r2, YX=1,Y=1
store 1, Xstore 1, Yload r2, Xload r2, YX=1,Y=1
store 1, Ystore 1, Xload r2, Yload r2, XX=1,Y=1
![Page 26: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/26.jpg)
Realistic Memory Models• Modern hardware features can interfere with store order:
– write buffer (or store buffer or write-behind buffer)– instruction reordering (out-of-order execution)– superscalar execution and pipelining
• Each CPU/core keeps its own data consistent, but how is it viewed by others?
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 27: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/27.jpg)
![Page 28: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/28.jpg)
Write-buffers and SMP• Stores go to write buffer to hide memory
latency– And cache invalidates
• Loads read from write buffer if possible
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
CPU0
Cache
Store C…
Store B…
Store A….
CPU 0store r1, Astore r2, Bstore r3, Cload r4, A
![Page 29: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/29.jpg)
CPU 1CPU 0
store r1, Astore r2, Bstore r3, C
Write-buffers and SMP• When the buffer eventually drains, what order
does CPU1 see CPU0’s memory updates?
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
What happens in our example?
CPU0
Cache
Store C…
Store B…
Store A….
![Page 30: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/30.jpg)
Total Store Ordering (e.g. x86)• Stores are guaranteed to occur in FIFO order
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
CPU0
Cache
Store C…
Store B…
Store A….
CPU 1 seesA=1B=2C=3
CPU 0store 1, Astore 2, Bstore 3, C
![Page 31: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/31.jpg)
Total Store Ordering (e.g. x86)• Stores are guaranteed to occur in FIFO order
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
CPU0
Cache
…Store mutex
…Store count
….
CPU 1 seescount updatedmutex = 0
/* counter++ */load r1, countadd r1, r1, 1store r1, counter
/* unlock(mutex) */store zero, mutex
![Page 32: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/32.jpg)
Total Store Ordering (e.g. x86)• Assume X = Y = 0 initially
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
What is the problem here?
CPU 0store 1, Xload r2, Y
CPU 1store 1, Yload r2, X
CPU0
Cache
…Store X
….
CPU1
Cache
…Store Y
….
![Page 33: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/33.jpg)
Total Store Ordering (e.g. x86)• Stores are buffered in write-buffer and
don’t appear on other CPU in time.• Can get X=0, Y=0!!!!• Loads can “appear” re-ordered with
preceding stores
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
load r2, Yload r2, Xstore 1, Xstore 1, Y
CPU 0store 1, Xload r2, Y
CPU 1store 1, Yload r2, X
CPU0
Cache
…Store X
….
CPU1
Cache
…Store Y
….
![Page 34: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/34.jpg)
Memory “fences” or “barriers”• The provide a “fence” between
instructions to prevent apparent re-ordering
• Effectively, they drain the local CPU’s write-buffer before proceeding.
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
CPU 0store 1, Xfenceload r2, Y
CPU 1store 1, Yfenceload r2, X
CPU0
Cache
…Store X
….
CPU1
Cache
…Store Y
….
![Page 35: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/35.jpg)
Total Store Ordering• Stores are guaranteed to occur in FIFO order• Atomic operations?
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
• Need hardware support, e.g.• atomic swap• test & set• load-linked + store-conditional
• Stall pipeline and drain (and bypass) write buffer• Ensures addr1 held exclusively
CPU 0ll r1, addr1sc r1, addr1
CPU 1ll r1, addr1sc r2, addr1
CPU
Cache
Store A…
Store B…
Store A….
![Page 36: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/36.jpg)
Partial Store Ordering (e.g. ARM MPcore)
• All stores go to write buffer• Loads read from write buffer if possible• Redundant stores are cancelled or merged
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
• Stores can appear to overtake (be re-ordered) other stores• Need to use memory barrier
CPU
Cache
Store A…
Store B…
Store A….
CPU 1 seesaddr2 = VALaddr1 = IDLE
CPU 0store BUSY, addr1store VAL, addr2store IDLE, addr1
![Page 37: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/37.jpg)
Partial Store Ordering (e.g. ARM MPcore)
• The barriers prevent preceding stores appearing after successive stores– Note: Reality is a little more complex (read barriers,
write barriers), but principle similar.
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
• Store to counter can overtake store to mutex• i.e. update move outside the lock
• Need to use memory barrier• Failure to do so will introduce subtle bugs:
• Critical section “leaking” outside the lock
load r1, counteradd r1, r1, 1store r1, counterbarrierstore zero, mutex
CPU
Cache
Store A…
Store B…
Store A….
![Page 38: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/38.jpg)
MP Hardware Take Away• Each core/cpu sees sequential execution of own
code• Other cores see execution affected by
– Store order and write buffers– Cache coherence model– Out-of-order execution
• Systems software needs understand:– Specific system (cache, coherence, etc..)– Synch mechanisms (barriers, test_n_set, load_linked
– store_cond).…to build cooperative, correct, and scalable parallel code
![Page 39: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/39.jpg)
MP Hardware Take Away• Existing sync primitives (e.g. locks) will have
appropriate fences/barriers in place– In practice, correctly synchronised code can ignore memory
model.• However, racey code, i.e. code that updates shared
memory outside a lock (e.g. lock free algorithms) must use fences/barriers.– You need a detailed understanding of the memory coherence
model.– Not easy, especially for partial store order (ARM).
![Page 40: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/40.jpg)
Memory ordering for various ArchitecturesType Alpha ARMv7 PA-RISC POWER SPARC
RMOSPARC PSO
SPARC TSO x86 x86
oostore AMD64 IA-64 zSeries
Loads reordered after loads
Y Y Y Y Y Y Y
Loads reordered after stores
Y Y Y Y Y Y Y
Stores reordered after stores
Y Y Y Y Y Y Y Y
Stores reordered after loads
Y Y Y Y Y Y Y Y Y Y Y Y
Atomic reordered with loads
Y Y Y Y Y
Atomic reordered with stores
Y Y Y Y Y Y
Dependent loads reordered
Y
Incoherent instruction cache pipeline
Y Y Y Y Y Y Y Y Y Y
![Page 41: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/41.jpg)
Concurrency Observations• Locking primitives require exclusive access to the “lock”
– Care required to avoid excessive bus/interconnect traffic
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 42: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/42.jpg)
Kernel Locking• Several CPUs can be executing kernel code
concurrently.• Need mutual exclusion on shared kernel data.• Issues:
– Lock implementation– Granularity of locking (i.e. parallelism)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 43: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/43.jpg)
Mutual Exclusion Techniques • Disabling interrupts (CLI — STI).
– Unsuitable for multiprocessor systems.• Spin locks.
– Busy-waiting wastes cycles.• Lock objects (locks, semaphores).
– Flag (or a particular state) indicates object is locked.– Manipulating lock requires mutual exclusion.
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 44: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/44.jpg)
Hardware Provided Locking Primitives
• int test_and_set(lock *);• int compare_and_swap(int c,
int v, lock *);• int exchange(int v, lock *)• int atomic_inc(lock *)
• v = load_linked(lock *) / boolstore_conditional(int, lock *)– LL/SC can be used to implement all of the above
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 45: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/45.jpg)
Spin locksvoid lock (volatile lock_t *l) {
while (test_and_set(l)) ;}void unlock (volatile lock_t *l) {
*l = 0;}• Busy waits. Good idea?
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 46: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/46.jpg)
Spin Lock Busy-waits Until Lock Is Released
• Stupid on uniprocessors, as nothing will change while spinning.– Should release (yield) CPU immediately.
• Maybe ok on SMPs: locker may execute on other CPU.– Minimal overhead (if contention low).– Still, should only spin for short time.
• Generally restrict spin locking to:– short critical sections,– unlikely to (or preferably can’t) be contended by the same CPU.– local contention can be prevented
o by designo by turning off interrupts
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 47: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/47.jpg)
Spinning versus Switching– Blocking and switching
o to another process takes time Save context and restore another Cache contains current process not new
» Adjusting the cache working set also takes time TLB is similar to cache
o Switching back when the lock is free encounters the same again
– Spinning wastes CPU time directly• Trade off
– If lock is held for less time than the overhead of switching to and back
It’s more efficient to spin
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 48: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/48.jpg)
Spinning versus Switching• The general approaches taken are
– Spin forever– Spin for some period of time, if the lock is not
acquired, block and switchoThe spin time can be Fixed (related to the switch overhead) Dynamic
» Based on previous observations of the lock acquisition time
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 49: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/49.jpg)
Interrupt Disabling• Assume no local contention by design, is disabling
interrupt important?
• Hint: What happens if a lock holder is preempted (e.g., at end of its timeslice)?
• All other processors spin until the lock holder is re-scheduled
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 50: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/50.jpg)
Alternative to spinning:Conditional Lock (TryLock)
bool cond_lock (volatile lock t *l) {if (test_and_set(l))
return FALSE; //couldn’t lockelse
return TRUE; //acquired lock}
• Can do useful work if fail to acquire lock.• But may not have much else to do.• Starvation: May never get lock!
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 51: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/51.jpg)
Another alternative to spinining.
void mutex lock (volatile lock t *l) {while (1) {
for (int i=0; i<MUTEX N; i++)if (!test and set(l))
return;yield();
}}
• Spins for limited time only– assumes enough for other CPU to exit critical section
• Useful if critical section is shorter than N iterations.• Starvation possible.
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 52: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/52.jpg)
Common Multiprocessor Spin Lockvoid mp_spinlock (volatile lock t *l) {
cli(); // prevent preemptionwhile (test and set(l)) ; // lock
}void mp unlock (volatile lock t *l) {
*l = 0;sti();
}
• Only good for short critical sections• Does not scale for large number of processors• Relies on bus-arbitrator for fairness• Not appropriate for user-level• Used in practice in small SMP systems
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 53: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/53.jpg)
Need a more systematic analysis
Thomas Anderson, “The Performance of Spin Lock Alternatives for Shared-Memory Multiprocessors”, IEEE Transactions on Parallel and Distributed Systems, Vol 1, No. 1, 1990
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 54: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/54.jpg)
Compares Simple Spinlocks• Test and Set
void lock (volatile lock_t *l) {while (test_and_set(l)) ;
}
• Test and Test and Set
void lock (volatile lock_t *l) {while (*l == BUSY || test_and_set(l)) ;
}
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 55: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/55.jpg)
test_and_test_and_set LOCK• Avoid bus traffic contention caused by test_and_set until it is likely to succeed • Normal read spins in cache• Can starve in pathological cases
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 56: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/56.jpg)
Benchmarkfor i = 1 .. 1,000,000 {
lock(l)crit_section()unlock()compute()
}
• Compute chosen from uniform random distribution of mean 5 times critical section
• Measure elapsed time on Sequent Symmetry (20 CPU 30386, coherent write-back invalidate caches)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 57: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/57.jpg)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 58: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/58.jpg)
Results• Test and set performs poorly once there is enough CPUs
to cause contention for lock– Expected
• Test and Test and Set performs better– Performance less than expected– Still significant contention on lock when CPUs notice release and
all attempt acquisition• Critical section performance degenerates
– Critical section requires bus traffic to modify shared structure– Lock holder competes with CPU that missed as they test and set
o lock holder is slower– Slower lock holder results in more contention
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 59: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/59.jpg)
Idea• Can inserting delays reduce bus traffic and
improve performance• Explore 2 dimensions
– Location of delayo Insert a delay after release prior to attempting acquireo Insert a delay after each memory reference
– Delay is static or dynamico Static – assign delay “slots” to processors
Issue: delay tuned for expected contention levelo Dynamic – use a back-off scheme to estimate contention
Similar to ethernet Degrades to static case in worst case.
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 60: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/60.jpg)
Examining Inserting Delays
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 61: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/61.jpg)
Queue Based Locking• Each processor inserts itself into a waiting queue
– It waits for the lock to free by spinning on its own separate cache line
– Lock holder frees the lock by “freeing” the next processors cache line.
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 62: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/62.jpg)
Results
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 63: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/63.jpg)
Results• Static backoff has higher overhead when backoff
is inappropriate• Dynamic backoff has higher overheads when
static delay is appropriate– as collisions are still required to tune the backoff time
• Queue is better when contention occurs, but has higher overhead when it does not.– Issue: Preemption of queued CPU blocks rest of
queue (worse than simple spin locks)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 64: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/64.jpg)
• John Mellor-Crummey and Michael Scott, “Algorithms for Scalable Synchronisation on Shared-Memory Multiprocessors”, ACM Transactions on Computer Systems, Vol. 9, No. 1, 1991
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 65: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/65.jpg)
MCS Locks• Each CPU enqueues its own private lock variable into a queue and
spins on it– No contention
• On lock release, the releaser unlocks the next lock in the queue– Only have bus contention on actual unlock– No starvation (order of lock acquisitions defined by the list)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 66: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/66.jpg)
MCS Lock• Requires
– compare_and_swap() – exchange()
oAlso called fetch_and_store()
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 67: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/67.jpg)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 68: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/68.jpg)
![Page 69: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/69.jpg)
Sample MCS code for ARM MPCore
void mcs_acquire(mcs_lock *L, mcs_qnode_ptr I) {
I->next = NULL;MEM_BARRIER;mcs_qnode_ptr pred = (mcs_qnode*) SWAP_PTR( L, (void *)I);if (pred == NULL) { /* lock was free */
MEM_BARRIER;return;
}I->waiting = 1; // word on which to spinMEM_BARRIER;pred->next = I; // make pred point to me
}
![Page 70: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/70.jpg)
Selected Benchmark• Compared
– test and test and set– Anderson’s array based queue– test and set with exponential back-off– MCS
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 71: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/71.jpg)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 72: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/72.jpg)
Confirmed Trade-off• Queue locks scale well but have higher overhead • Spin Locks have low overhead but don’t scale well• What do we use?
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 73: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/73.jpg)
• Beng-Hong Lim and Anant Agarwal, “Reactive Synchronization Algorithms for Multiprocessors”, ASPLOS VI, 1994
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 74: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/74.jpg)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 75: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/75.jpg)
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 76: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/76.jpg)
Idea• Can we dynamically switch locking methods to
suit the current contention level???
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 77: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/77.jpg)
Issues• How do we determine which protocol to use?
– Must not add significant cost• How do we correctly and efficiently switch protocols?• How do we determine when to switch protocols?
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 78: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/78.jpg)
Protocol Selection• Keep a “hint”• Ensure both TTS and MCS lock a never free at the same
time– Only correct selection will get the lock– Choosing the wrong lock with result in retry which can get it right
next time– Assumption: Lock mode changes infrequently
o hint cached read-onlyo infrequent protocol mismatch retries
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 79: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/79.jpg)
Changing Protocol• Only lock holder can switch to avoid race conditions
– It chooses which lock to free, TTS or MCS.
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 80: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/80.jpg)
When to change protocol• Use threshold scheme
– Repeated acquisition failures will switch mode to queue– Repeated immediate acquisition will switch mode to TTS
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 81: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/81.jpg)
Results
© Kevin Elphinstone. Distributed under Creative Commons Attribution License
![Page 82: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/82.jpg)
The multicore evolution and operating systems
Frans Kaashoek
Joint work with: Silas Boyd-Wickizer, Austin T. Clements, Yandong Mao, Aleksey Pesterev, Robert Morris, and Nickolai
Zeldovich
MIT
![Page 83: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/83.jpg)
• Non-scalable locks are dangerous.Silas Boyd-Wickizer, M. Frans Kaashoek, Robert Morris, and Nickolai Zeldovich. In the Proceedings of the Linux Symposium, Ottawa, Canada, July 2012.
![Page 84: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/84.jpg)
How well does Linux scale?
● Experiment: ● Linux 2.6.35-rc5 (relatively old, but problems are representative of issues in recent kernels too) ● Select a few inherent parallel system applications ● Measure throughput on different # of cores ● Use tmpfs to avoid disk bottlenecks
● Insight 1: Short critical sections can lead to sharp performance collapse
![Page 85: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/85.jpg)
Off-the-shelf 48-core server (AMD) DRAM DRAM DRAM DRAM
DRAM DRAM DRAM DRAM
● Cache-coherent and non-uniform access ● An approximation of a future 48-core chip
![Page 86: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/86.jpg)
Poor scaling on stock Linux kernel 48
perfect scaling 44
40
36
32
28
24
20
16
12
8 terrible scaling
4
0 memcached PostgreSQL Psearchy
Exim Apache gmake Metis
Y-axis: (throughput with 48 cores) / (throughput with one core)
![Page 87: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/87.jpg)
Exim on stock Linux: collapse 12000
Throughput
10000
8000
6000
4000
2000
0 1 4 8 12 16 20 24 28 32 36 40 44 48
Cores
![Page 88: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/88.jpg)
Exim on stock Linux: collapse 12000
Throughput
10000
8000
6000
4000
2000
0 1 4 8 12 16 20 24 28 32 36 40 44 48
Cores
![Page 89: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/89.jpg)
Exim on stock Linux: collapse 12000 15
Throughput Kernel time
10000 12
8000
9
6000
6
4000
3 2000
0 01 4 8 12 16 20 24 28 32 36 40 44 48
Cores
![Page 90: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/90.jpg)
Oprofile shows an obvious problem
40 cores: 10000 msg/sec
48 cores: 4000 msg/sec
samples % app name 2616 7.3522 vmlinux 2329 6.5456 vmlinux 2197 6.1746 vmlinux 1488 4.1820 vmlinux 1348 3.7885 vmlinux 1182 3.3220 vmlinux 966 2.7149 vmlinux
samples % app name 13515 34.8657 vmlinux 2002 5.1647 vmlinux 1661 4.2850 vmlinux 1497 3.8619 vmlinux 1026 2.6469 vmlinux 914 2.3579 vmlinux 896 2.3115 vmlinux
symbol name radix_tree_lookup_slot unmap_vmas filemap_fault __do_fault copy_page_c unlock_page page_fault
symbol name lookup_mnt radix_tree_lookup_slot filemap_fault unmap_vmas __do_fault atomic_dec unlock_page
![Page 91: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/91.jpg)
Oprofile shows an obvious problem
40 cores: 10000 msg/sec
48 cores: 4000 msg/sec
samples % app name 2616 7.3522 vmlinux 2329 6.5456 vmlinux 2197 6.1746 vmlinux 1488 4.1820 vmlinux 1348 3.7885 vmlinux 1182 3.3220 vmlinux 966 2.7149 vmlinux
samples % app name 13515 34.8657 vmlinux 2002 5.1647 vmlinux 1661 4.2850 vmlinux 1497 3.8619 vmlinux 1026 2.6469 vmlinux 914 2.3579 vmlinux 896 2.3115 vmlinux
symbol name radix_tree_lookup_slot unmap_vmas filemap_fault __do_fault copy_page_c unlock_page page_fault
symbol name lookup_mnt radix_tree_lookup_slot filemap_fault unmap_vmas __do_fault atomic_dec unlock_page
![Page 92: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/92.jpg)
Oprofile shows an obvious problem samples % app name symbol name2616 7.3522 vmlinux radix_tree_lookup_slot2329 6.5456 vmlinux unmap_vmas
40 cores: 2197 6.1746 vmlinux filemap_fault10000 msg/sec 1488 4.1820 vmlinux __do_fault
1348 3.7885 vmlinux copy_page_c1182 3.3220 vmlinux unlock_page966 2.7149 vmlinux page_fault
samples % app name symbol name13515 34.8657 vmlinux lookup_mnt
48 cores: 4000 msg/sec
2002 5.1647 vmlinux 1661 4.2850 vmlinux 1497 3.8619 vmlinux 1026 2.6469 vmlinux 914 2.3579 vmlinux 896 2.3115 vmlinux
radix_tree_lookup_slot filemap_fault unmap_vmas __do_fault atomic_dec unlock_page
![Page 93: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/93.jpg)
Bottleneck: reading mount table ● Delivering an email calls sys_open ● sys_open calls
struct vfsmount *lookup_mnt(struct path *path) {
struct vfsmount *mnt; spin_lock(&vfsmount_lock); mnt = hash_get(mnts, path); spin_unlock(&vfsmount_lock); return mnt;
}
![Page 94: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/94.jpg)
Bottleneck: reading mount table
● sys_open calls: struct vfsmount *lookup_mnt(struct path *path) {
struct vfsmount *mnt; spin_lock(&vfsmount_lock); mnt = hash_get(mnts, path); spin_unlock(&vfsmount_lock); return mnt;
}
![Page 95: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/95.jpg)
Bottleneck: reading mount table
● sys_open calls: struct vfsmount *lookup_mnt(struct path *path) {
struct vfsmount *mnt; spin_lock(&vfsmount_lock); mnt = hash_get(mnts, path); spin_unlock(&vfsmount_lock); return mnt;
}
Serial section is short. Why does it cause a scalability bottleneck?
![Page 96: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/96.jpg)
What causes the sharp performance collapse?
● Linux uses ticket spin locks, which are non-scalable
● So we should expect collapse [Anderson 90]
● But why so sudden, and so sharp, for a short section?
● Is spin lock/unlock implemented incorrectly? ● Is hardware cache-coherence protocol at fault?
![Page 97: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/97.jpg)
Scalability collapse caused by non-scalable locks [Anderson 90]
void spin_lock(spinlock_t *lock) {
t = atomic_inc(lock->next_ticket); while (t != lock->current_ticket)
; /* Spin */ }
void spin_unlock(spinlock_t *lock) {
lock->current_ticket++; }
struct spinlock_t { int current_ticket; int next_ticket;
}
![Page 98: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/98.jpg)
Scalability collapse caused by non-scalable locks [Anderson 90]
void spin_lock(spinlock_t *lock) {
t = atomic_inc(lock->next_ticket); while (t != lock->current_ticket)
; /* Spin */ }
void spin_unlock(spinlock_t *lock) {
lock->current_ticket++; }
struct spinlock_t { int current_ticket; int next_ticket;
}
![Page 99: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/99.jpg)
Scalability collapse caused by non-scalable locks [Anderson 90]
void spin_lock(spinlock_t *lock) {
t = atomic_inc(lock->next_ticket); while (t != lock->current_ticket)
; /* Spin */ }
void spin_unlock(spinlock_t *lock) {
lock->current_ticket++; }
struct spinlock_t { int current_ticket; int next_ticket;
}
![Page 100: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/100.jpg)
Scalability collapse caused by non-scalable locks [Anderson 90]
void spin_lock(spinlock_t *lock) {
t = atomic_inc(lock->next_ticket); while (t != lock->current_ticket)
; /* Spin */ }
void spin_unlock(spinlock_t *lock) {
lock->current_ticket++; }
struct spinlock_t { int current_ticket; int next_ticket;
}
![Page 101: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/101.jpg)
Scalability collapse caused by non-scalable locks [Anderson 90]
void spin_lock(spinlock_t *lock) {
t = atomic_inc(lock->next_ticket); while (t != lock->current_ticket)
; /* Spin */ }
void spin_unlock(spinlock_t *lock) {
lock->current_ticket++; }
struct spinlock_t { int current_ticket; int next_ticket;
}
![Page 102: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/102.jpg)
Scalability collapse caused by non-scalable locks [Anderson 90]
void spin_lock(spinlock_t *lock) {
t = atomic_inc(lock->next_ticket); while (t != lock->current_ticket)
; /* Spin */ }
500 cycles
void spin_unlock(spinlock_t *lock) {
lock->current_ticket++; }
struct spinlock_t { int current_ticket; int next_ticket;
}
![Page 103: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/103.jpg)
Scalability collapse caused by non-scalable locks [Anderson 90]
void spin_lock(spinlock_t *lock) {
t = atomic_inc(lock->next_ticket); while (t != lock->current_ticket)
; /* Spin */ }
void spin_unlock(spinlock_t *lock) {
lock->current_ticket++; }
struct spinlock_t { int current_ticket; int next_ticket;
}
![Page 104: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/104.jpg)
Scalability collapse caused by non-scalable locks [Anderson 90]
void spin_lock(spinlock_t *lock) {
t = atomic_inc(lock->next_ticket); while (t != lock->current_ticket)
; /* Spin */ }
void spin_unlock(spinlock_t *lock) {
lock->current_ticket++; }
struct spinlock_t { int current_ticket; int next_ticket;
}
![Page 105: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/105.jpg)
Scalability collapse caused by non-scalable locks [Anderson 90]
void spin_lock(spinlock_t *lock) {
t = atomic_inc(lock->next_ticket); while (t != lock->current_ticket)
; /* Spin */ }
void spin_unlock(spinlock_t *lock) {
lock->current_ticket++; }
struct spinlock_t { int current_ticket; int next_ticket;
}
Previous lock holder notifies next lock holder after
sending out N/2 replies
![Page 106: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/106.jpg)
Why collapse with short sections?
● Arrival rate is proportional to # non-waiting cores ● Service time is proportional to # cores waiting (k)
● As k increases, waiting time goes up ● As waiting time goes up, k increases
● System gets stuck in states with many waiting cores
![Page 107: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/107.jpg)
Short sections result in collapse
● Experiment: 2% of time spent in critical section ● Critical sections become “longer” with more cores ● Lesson: non-scalable locks fine for long sections
![Page 108: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/108.jpg)
Avoiding lock collapse ● Unscalable locks are fine for long sections ● Unscalable locks collapse for short sections
● Sudden sharp collapse due to “snowball” effect ● Scalable locks avoid collapse altogether
● But requires interface change
![Page 109: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/109.jpg)
Scalable lock scalability
● It doesn't matter much which one ● But all slower in terms of latency
![Page 110: SMP, Multicore, Memory Ordering & Locking](https://reader030.fdocuments.in/reader030/viewer/2022013000/61c97a0a82f756392355ab03/html5/thumbnails/110.jpg)
Avoiding lock collapse is not enough to scale
● “Scalable” locks don't make the kernel scalable ● Main benefit is avoiding collapse: total throughput will not be lower with more cores ● But, usually want throughput to keep increasing with more cores