Transcript lecture31-dec10

Operating Systems
CSE 411
Multi-processor Operating Systems
Dec. 8 2006 - Lecture 30
Instructor: Bhuvan Urgaonkar
Grading Concerns
• Exam 1 distribution
–
–
–
–
–
–
80+
70+
60+
50+
40+
40-
• Exam 2 distribution
–
–
–
–
–
–
80+ : 4
70+ : 5
60+ : 9
50+ : 18
40+ : 17
40- : 7
A, AB+, B
B-, C+
C, C-, …
• Performance on homeworks/quizzes/projects has been much better
than on exams
• Project 3 will be made easier
• More time for Project 2
• Last quiz to replace homework 5
– Will be long and comprehensive
– Will help you prepare for the final exam
Multi-processor Computers
• Asymmetric
– Master processor runs the kernel
– Other processors run user code
• Symmetric
– Each server is self-scheduling
• Single run queue or per -processor run queue
– Scheduling and synchronization more complex
Overview of Symmetric Multiprocessor
(SMP)
CPU 0
CPU 1
HW cache
HW cache
RAM
• More CPUs: more processing power but more contention for the bus
Hardware features of Modern SMPs
• Common memory
– All CPUs connected to a common bus
– Hardware circuit called memory arbiter inserted between the bus and each
RAM chip
• Serializes accesses to memory
• Actually there is an arbiter even in uni-processors
– Why: Remember DMA?
• Hardware support for cache synchronization
– Contents of the cache and memory maintain their consistency at the
hardware level just as in a uni-processor
– Cache snooping: Whenever a CPU modifies its cache, it must check if the
same data is contained in another cache, and if so, notify it of the update
• Note: Implemented in hardware, not of concern to the kernel
Hardware features of Modern SMPs
(contd.)
• Distributed Interrupt Processing
– Being able to deliver interrupt to any CPU crucial to exploit the parallelism
of the SMP architecture
– Special hardware for routing interrupts to the right processors
– Interprocessor interrupts
• Used fot TLB consistency among other things
OS Issues for SMP
• How should the CPU scheduler work?
• How should synchronization work?
CPU Scheduling in a Multi-processor
• Kernel and user code can now run on N processors
– A multi-threaded process may span multiple CPUs
• Uni-processor: Scheduler picks a process to run
• Multi-processor: Scheduler picks a process and the CPU on
which it will run
• A process may move from one CPU to another during its
lifetime
• Two main factors affecting scheduler design
– Cache affinity: Would like to run a process on the same CPU
– Load balancing: Would like to keep all CPUs equally busy
– These are often at odds with each other: Why?
Proportional Fair Scheduling
in SMPs
• What will happen if we have two threads
with weights 1 and 10 and a Lottery
scheduler on a dual-processor?
• Not all combinations of weights are feasible!
• Given k threads with weights w1, …, wk, and
p processors, can you think of a feasibility
criterion?
Symmetric Multi-threading
• Idea: Create multiple virtual processors on a physical processor
–
–
–
–
Illusion provided by hardware
Called hyper-threading in Intel machines
The OS sees the machine as an SMP
Each virtual processor has its own set of registers and interrupt handling,
cache is shared
• Why: Possibility of better parallelism, better utilization
• How does it concern the OS?
– From a correctness point of view: it does not
– From an efficiency point of view: the scheduler could exploit it to do better
load balancing
Synchronization in SMPs
Synchronization building blocks:
Atomic Instructions
• An atomic instruction for a uni-processor would not be atomic
on an SMP unless special care is taken
– In particular, any atomic instruction needs to write to a memoy location
– Recall TestAndSet, Swap
• Need special support from the hardware
– Hardware needs to ensure that when a CPU writes to a memory
location as part of an atomic instruction, another CPU can not write
the same memory location till the first CPU is finished with its write
• Given above hardware support, OS support for
synchronization of user processes same as in uni-processors
– Semaphores, monitors
• Kernel synchronization raises some special considerations
– Both in uni- and multi-processors
Kernel synchronization
Interlude: Back to the past:
A little background (mostly revision)
• Recall: A kernel is a “server” that answers requests issued in two
possible ways
– A process causes an exception (E.g., page fault, system call)
– An external device sends an interrupt
• Definition: Kernel Control Path
– The set of instructions executed in the kernel mode to handle a kernel
request
– Similar to a process, except much more rudimentary
• No descriptor of any kind
– Most modern kernels are “re-entrant” => Multiple KCPs may be executing
simultaneously
• Synchronization problems can occur if two KCPs update the same data
– How to synchronize KCP access to shared data/resources?
How to synchronize KCP
access to shared data?
• Can use semaphores or monitors
• Not the best solution in multi-processors
• Consider two KCPs running on different CPUs
– If the time to update a shared data structure is very short, then semaphores
may be an overkill
• Solution: Spin Locks
– Do busy wait instead of getting blocked!
– The kernel programmer must decide when to use spin locks versus
semaphores
– Spin locks are useless in uni-processors: Why?
Two other mechanisms for easy to
achieve kernel synchronization
•
Non-preemptible kernel design
– A KCP can not be pre-empted by another one
• Useful only certain KCPs, such as those that have no synch. issues with interrupt handlers
– Not enough for multi-processors since multiple CPUs can concurrently access the
same data structure
– Adopted by many Oses including versions of Linux upto 2.4
– Linux 2.6 is pre-emptible: faster dispatch times for user processes
•
Interrupt disabling
– Disable all hardware interrupts before entering a critical section and re-enable them
right after leaving it
– Works for uni-processor in certain situations
• The critical section should not incur an exception whose handler has synchronization issues
with it
• The CS should not get blocked
–
What if the CS incurs a page fault?
– Does not work for multi-processors
• Interrupts must be