Black, D.L. Scheduling and Resource Management Techniques for Multiprocessors. PhD thesis, Department of Computer Science, Carnegie Mellon University, July, 1990.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....context switching, we felt that further enhancements could be made. The old implementation is very aggressive about blocking. Spinning before blocking has been shown to be superior to just spinning or just blocking [12] Mach 3. 0 has better scheduling primitives for synchronization then mach msg( [5]. Finally, the old Mach 3.0 code is complex and hard to maintain. We created the new implementation to address our concerns with the shortcomings of the old implementation. A continuation is an object that fully describes a future computation. A continuation can be built, invoked, and passed ....

....waiting for a C Thread to appear that can be run. If this occurs, the idle thread attempts to acquire the run lock, dequeue a C Thread, and switch to it. After spinning for the maximum number of iterations and not having a C Thread appear, the idle thread depresses its priority and begins again [5]. This priority depression yields the processor to any runnable kernel thread. When there are no runnable kernel threads except the depressed thread, the system automatically aborts the priority depression and resumes the idle thread. We call this depressed priority spinning or Busy Waiting. ....

Black, D.L. Scheduling and Resource Management Techniques for Multiprocessors. PhD thesis, Department of Computer Science, Carnegie Mellon University, July, 1990.

....other kernels examined attempts to be intelligent about sharing of such queues across a large system. Periodic scan of all threads in the run queues is performed to avoid starvation, although this is done more eciently by having each thread maintain its own usage information. This is described in [6] with a method that appears complex and tailored for a speci c scheduler design. It provides load balancing in the kernel which, while being ecient if well implemented, may restrict user implementation of alternate scheduling procedures. This method would be of concern particularly for the NUMA ....

D. L. Black. Scheduling and Resource Management Techniques for Multiprocessors. School of Computer Science, Carnegie Mellon University, Pittsburgh, 1999.

....they require hardware support not available in most existing machines. The algorithms require a reference counter for each frame processor combination. Such hardware requirements present a problem of scalability as the number of processors grows. In addition, as Black acknowledges in his thesis [Bla90] the work presented in Chapter 6 of this dissertation raises questions about the cost effectiveness of the additional hardware support required, since simpler policies perform as well as approximations to the competitive policy. However, the PLUS system currently being developed at CMU [BR90b, ....

D. Black. Scheduling and Resource Management Techniques for Multiprocessors. PhD thesis, Carnegie-Mellon University, July 1990.

....microsecond counter available on the machine. The Mach operating system provides a system call to map this counter into the address space of a task, so that taking a timestamp is simply a user space memory reference that takes approximately 3 #s. Mach also provides a processor allocation facility [7] that grants users exclusive access to a set of processors for a limited amount of time. Access to this facility is mediated by a privileged server. The maximum time that can be requested from this server is 15 minutes, and no more than threequarters of the system (12 processors for our ....

BLACK,D.L. Scheduling and Resource Management Techniques for Multiprocessors. PhD thesis, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, July 1990. 54

....is referred to as information lag in the literature and methods to compensate for this are discussed in 3.2.4. 2.5. A Related Problem: Job Load Balancing A substantial amount of research has been conducted on load balancing in general distributed systems [CK88] and distributed operating systems [Bla90b, Bla90a]. However, these studies usually assumed that each job can be processed in its entirety by any node and that each job only requires CPU and memory resources. In networks of workstations, the primary concerns are the utilization of idling workstations, and the migration of long jobs away from ....

D. L. Black. "Scheduling and Resource Management Techniques for Multiprocessors". Research Report (Ph.D. thesis) CMU-CS-90-152, CMU, 1990.

....the lock ownership information in the registration module. An attempt to re acquire the same lock is easily detected because a lock object knows the identity of its owner. Recursive locks are more expensive than the normal locks because each requesting 1 recursive locks are implemented in MACH [Bla90b] kernel to eliminate the lock reentrancy problem. thread performs an extra memory write at registration time. Simple conditional locks result when the timeout parameter is set accordingly. A thread waiting for such a lock returns unsuccessfully if it cannot acquire the lock in a specified time. ....

David. L. Black. Scheduling and Resource Management Techniques for Multiprocessors. PhD thesis, School of Computer Science, Carnegie Mellon University, July 1990. Techreport CMUCS -90-152.

....KSR 1 machine) the receiving processor is interrupted after the sender has put the arriving message into the receiver s message list. The slot list is updated after the reschedule. There are two interrupts associated with each slot: at the leading edge of the slot, the processor is handed off [4] to the thread that will run that particular event, while at the trailing edge of the slot, the processor is handed off to the dispatcher thread 7 , which later hands the processor off to a particular event at the next leading edge of a slot. Unix timers (setitimer( and signal( and ....

D. L. Black. Scheduling and resource management techniques for multiprocessors. (CMU-CS-90-152), July 1990.

....communicating partners of a process. In a multi threaded system like ODOS the communicating partners are at least the threads of one task. All these threads use the same virtual memory. It is therefore clear that they can run on different processors of one shared memory multi processor [Bla90b, Bla90a, Wen87] They can also migrate from one processor to another. In the shared memory system they can access memory pages randomly. A difficult situation is when threads are spread out onto different sites and not all memory pages are in the local memory. Threads which operate on the same memory ....

David L. Black. Scheduling and Resource Management Techniques for Multiprocessors. Master's thesis, School of Computer Science, Carnegie Mellon University, 1990.

....the past CPU time has been obtained, the smaller the priority depression it entails. In addition, the priorities of jobs may have fixed base components, yielding the opportunity to give different jobs different levels of service. It is this type of scheduling that is used in UNIX, and also in Mach [2, 3], which latter is the basis of the operating system of the Open Software Foundation (OSF) 1 . In the continuous time multiprocessor scheduling policy we study, the priority of a job is equal to the sum of its fixed base priority and a time dependent component. The latter part, the accumulated ....

....well; however, after adapting the model to include the somewhat different way in which Mach effectuates decay, the measurements match the model very well indeed. Our analysis extends the analysis by J.L. Hellerstein [10] of scheduling in UNIX SystemV uniprocessors and the short treatment by Black [3] of Mach multiprocessors to a unified treatment of scheduling in uniprocessors and multiprocessors running UNIX System V (up to Release 4) 4.3BSD UNIX, 4.4BSD UNIX, or Mach. Neither of these authors proves that the decay usage scheduling policy reaches a steady state, nor do they deal with the ....

[Article contains additional citation context not shown here]

D.L. Black, Scheduling and Resource Management Techniques for Multiprocessors, Report CMU-CS-90-152, Carnegie Mellon University, 1990.

....is an area of active research. Bolosky, Scott, and Fitzgerald [7] and Cox and Fowler [10] demonstrated specific solutions implemented on the IBM Ace and BBN Butterfly Plus multiprocessors, respectively. Black et al. proposed provably competitive algorithms for page migration and replication in [4, 5, 6]. Scheurich and Dubois proposed page migration algorithms based on page pivoting in [26] Bolosky et al. conducted a trace based simulation study of the effects of architectural features on the performance of several migration and replication policies in [8] and Ramanathan and Ni conducted a ....

D. Black. Scheduling and Resource Management Techniques for Multiprocessors. PhD thesis, Carnegie-Mellon University, July 1990.

....the processors. The user is generally not aware of the multiprocessor nature of the machine: the parallelized kernel offers a single system image. For those users who need to master the parallelism, for example scientific computing applications, new scheduling techniques (e.g. gang scheduling [Black90] ) have been designed. Unfortunately, these systems currently suffer from scalability problems. The I O and memory buses become the bottlenecks of the system, which must be balanced by costly memory (caching) architectures. From the software point of view, the ability to support a large number of ....

David L. Black, "Scheduling and Resource Management Techniques for Multiprocessors, " in , Carnegie Mellon University, Pittsburgh, PA, (July, 1990), pp. 111. CS/EX91 -78

....is an area of active research. Bolosky, Scott, and Fitzgerald [5] and Cox and Fowler [8] demonstrated specific solutions implemented on the IBM Ace and BBN Butterfly Plus multiprocessors, respectively. Black et al. proposed provably competitive algorithms for page migration and replication in [2, 3, 4], Scheurich and Dubois proposed page migration algorithms based on page pivoting in [19] and Bolosky et al. conducted a trace based simulation study of the effects of architectural features on the performance of four migration and replication policies in [6] We have also investigated OS level ....

D. Black. Scheduling and Resource Management Techniques for Multiprocessors. PhD thesis, Carnegie-Mellon University, July 1990.

....consistency among multiple copies of a page when writes occur. Hence, programmers or operating systems restrict writable pages to a single copy. This gives rise to the page placement problem, which concerns the decision as to which local memory should contain the single page copy. Black et al. [35] refer to this problem as the migration problem. A similar problem is observed for read only pages. Specifically, the replication problem is concerned with determining the set of local memories that should contain copies of a page. Here, the assumption is made that the set of local memories that ....

....(network message servers) to extend the IPC across a network [100] On top of the general message primitives, Mach implements various flavors of communication including server client remote procedure calls, distributed object oriented programming, and streams. 4.3. 3 Scheduling The Mach scheduler [36, 34, 35] consists of two parts, one responsible for processor allocation, the other responsible for scheduling threads on individual processors. Processor Allocation. A user level server performs processor allocation, using the mechanisms provided by the underlying Mach kernel. The processor allocation ....

[Article contains additional citation context not shown here]

David. L. Black. Scheduling and Resource Management Techniques for Multiprocessors. PhD thesis, School of Computer Science, Carnegie Mellon University, July 1990. Techreport CMU-CS-90-152.

....KSR 1 machine) the receiving processor is interrupted after the sender has put the arriving message into the receiver s message list. The slot list is updated after the reschedule. There are two interrupts associated with each slot: at the leading edge of the slot, the processor is handed off [3] to the thread that will run that particular event, while at the trailing edge of the slot, the processor is handed off to the dispatcher thread 7 , which later hands the processor off to a particular event at the next leading edge of a slot. Unix timers (setitimer( and signal( and ....

David L. Black. Scheduling and resource management techniques for multiprocessors. (CMU-CS-90152) , July 1990.

....global and local scheduling as suggested by Krueger [Krue87] Krueger has proved that not every local scheduling algorithm suits the need of the distributed one and vice versa. In traditional systems it was hard to access and modify local schedulers, while Mach may be suitable for such experiments [Blac90, Blac90a]. Another area is relation between fine grained and coarse grained LD. It would be interesting to find out if systems like Amber [Chas89] could profit from the experience gained in our TM scheme. It actually deals with coarse grained support for fine grain objects. We may use experience gained by ....

Black, D. (1990) Scheduling and Resource Management Techniques for Multiprocessors. Technical Report CMU-CS-90-152, PhD Thesis.

....into the thread library. In conjunction with other mechanisms, to bind an application exclusively to a set of processors, interfering kernel scheduling can be avoided completely. This can be a job or gang scheduling strategy [4] 5] a processor reservation mechanism like the MACH CPU Server [6] or some other mechanisms for processor scheduling between applications integrated in the Sleeping Threads. 5 To describe the Sleeping Threads mechanism we need to introduce some new thread states. To implement the Sleeping Threads we need three additional kernel thread states: parked ....

D. J. Black, "Scheduling and Resource Management Techniques for Multiprocessors", Carnegie Mellon University, Technical Report CMU-CS-90-152, 1990

....system. In general, real time systems implement a simple fixed priority ordering [Hildebrand 92] Conventional operating systems, however, support either a single scheduling policy to arbitrate among competing activities, or multiple policies that promote fairness but favor interactive activities [Black 90] Some systems provide multiple scheduling policies but only from among a few fixed policies set at kernel build time [Tokuda et al. 90] Flexible, application specific scheduling, though, has been shown to provide critical performance benefits for both time constrained and non real time ....

Black, D. L. Scheduling and Resource Management Techniques for Multiprocessors. PhD dissertation, Carnegie Mellon University, July 1990.

....game, which runs only on the Toshiba. The table shows that about 99 of all control transfers use continuations and take advantage of stack discarding. The most frequent operations are message receive and exception handling. The other operations are page fault handling, voluntary rescheduling [Black 90a] involuntary preemptions, and blocking by internal kernel threads. The remaining blocking operations (which do not use continuations) occur during kernel mode page faults, memory allocation, and lock acquisition. Because generating a continuation for these cases is difficult, MK40, using the ....

Black, D. L. Scheduling and Resource Management Techniques for Multiprocessors. PhD dissertation, School of Computer Science, Carnegie Mellon University, July 1990.

....of the sender LP: which just schedules recovery events (which should have smaller execution time than the rejected event) on the LP that rejected the original event. Thus, there were two interrupts associated with each slot: at the leading edge of the slot, the processor is handed off [Bla90] to the thread that will run that particular event, while at the trailing edge of the slot, the processor is handed off to the dispatcher thread which later hands the processor off to a particular event at the next leading edge of a slot. Unix timers (setitimer( and signal( and context ....

David. L. Black. Scheduling and Resource Management Techniques for Multiprocessors. PhD thesis, School of Computer Science, Carnegie Mellon University, July 1990. Techreport CMU-CS-90-152.

....individual applications needs as well as to utilize the system s processor resources more efficiently. 2. 1 Related Work One simple approach to providing real time support in systems with traditional timesharing schedulers, which has been adopted by many commonly used systems such as Unix, Mach[1, 4], and Windows NT[25] and has even become part of the POSIX standard[17] is support for fixedpriority threads. Although these systems generally still use conventional priority based timesharing schedulers, they allow real time applications to disable the normal dynamic priority adjustments on ....

....possible CPU accounting mechanisms, each with different cost benefit tradeoffs. The CPU inheritance scheduling framework allows a variety of accounting policies to be implemented by scheduler threads. There are two well known approaches to CPU usage accounting: statistical and time stamp based[4]. With statistical accounting, the scheduler wakes up on every clock tick, checks the currently running thread, and charges the entire time quantum to that thread. This method is quite efficient, since the scheduler generally wakes up on every clock tick anyway; however, it provides limited ....

D. L. Black. Scheduling and Resource Management Techniques for Multiprocessors. PhD thesis, Carnegie Mellon University, July 1990.

....tracking blocked threads instead of relying on the event mechanism to do so. Such an implementation could block threads on event zero (the null event) from which only a clear wait can awaken them. Further details on these routines and the internal functioning of Mach s scheduler can be found in [4]. 7 Locks and Interrupts The interaction of locks and interrupts causes three problems: 1. The interrupt code may try to acquire a lock held by the code that was interrupted. Interrupt routines lack the thread context required to sleep, and are therefore forbidden from acquiring sleep locks. 2. ....

David L. Black, Scheduling and Resource Management Tech- niques for Multiprocessors, PhD thesis, Carnegie Mellon University, School of Computer Science, Pittsburgh, PA, (July, 1990), Available as Technical Report CMU-CS-90-152.

No context found.

D. L. Black, Scheduling and Resource Management Techniques for Multiprocessors, PhD thesis, Carnegie Mellon University, July 1990.

No context found.

David L. Black. "Scheduling and Resource Management Techniques for Multiprocessors", PhD Thesis, CMU-CS-90-152, Carnegie Mellon University, July 1990.

No context found.

David L. Black. Scheduling and Resource Management Techniques for Multiprocessors. Technical Report CMU-CS-90/152, The school of Computer Science, Carnegie Mellon University, July 1990.

No context found.

Black, D. L. Scheduling and Resource Management Techniques for Multiprocessors. Ph.D. dissertation, Carnegie Mellon University, July 1990.

First 50 documents

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC