diff options
author | venkita <venkita@ae88bc3d-4319-0410-8dbf-d08b4c9d3795> | 2003-10-12 17:03:25 +0000 |
---|---|---|
committer | venkita <venkita@ae88bc3d-4319-0410-8dbf-d08b4c9d3795> | 2003-10-12 17:03:25 +0000 |
commit | b838a60ec02db29469ba6ef4a0312667f3cd2d98 (patch) | |
tree | 250faa00ab93c06c86166a1122799018bd603a45 /Kokyu | |
parent | d14ee899787ae1f04ff5e0e233a5aaafbe7486df (diff) | |
download | ATCD-b838a60ec02db29469ba6ef4a0312667f3cd2d98.tar.gz |
ChangeLogTag: Sun Oct 12 11:48:02 2003 Venkita Subramonian <venkita@cs.wustl.edu>
Diffstat (limited to 'Kokyu')
-rw-r--r-- | Kokyu/Kokyu.html | 416 | ||||
-rw-r--r-- | Kokyu/KokyuEC.jpg | bin | 0 -> 58534 bytes | |||
-rw-r--r-- | Kokyu/kokyu1.jpg | bin | 0 -> 86410 bytes | |||
-rw-r--r-- | Kokyu/kokyu2.jpg | bin | 0 -> 82664 bytes |
4 files changed, 416 insertions, 0 deletions
diff --git a/Kokyu/Kokyu.html b/Kokyu/Kokyu.html new file mode 100644 index 00000000000..55c8016cd1c --- /dev/null +++ b/Kokyu/Kokyu.html @@ -0,0 +1,416 @@ +<!-- $Id$ --> +<!doctype html public "-//w3c//dtd html 4.0 transitional//en"> +<html> +<head> + <meta http-equiv="Content-Type" content="text/html; charset=iso-8859-1"> + <meta name="Author" content="Venkita Subramonian"> + <meta name="GENERATOR" content="Mozilla/4.79 [en] (Windows NT 5.0; U) [Netscape]"> + <title>Kokyu</title> +</head> +<body> + +<center> +<h2> +<b><font size=+2>Kokyu - A middleware framework for flexible scheduling +and dispatching</font></b></h2></center> +<a href="#Introduction">Introduction</a> +<br><a href="#SchedFramework">Strategized Scheduling framework</a> +<br><a href="#FlexDispatch">Flexible Dispatching Framework</a> +<br><a href="#KokyuEC">Use of Kokyu within the TAO Real-time Event Channel(RTEC)</a> +<br><a href="#ConfigKokyuEC">Configuration of RTEC to use Kokyu dispatching</a> +<br><a href="#KokyuDSRTCORBA">Use of Kokyu within the Dynamic Scheduling +Real-time CORBA (DSRTCORBA) schedulers</a> +<br><a href="#newDSRTSched">How to write a new DSRT scheduler using Kokyu</a> +<br><a href="#DSRTCORBAvsRTEC">Kokyu DSRTCORBA vs Kokyu RTEC</a> +<br><a href="#Status">Current status</a> +<br><a href="#Future">Future work</a> +<br><a href="#Papers">Papers on Kokyu</a> +<br> +<h3> +<a NAME="Introduction"></a>Introduction</h3> +Kokyu is a portable middleware scheduling framework designed to provide +flexible scheduling and dispatching services within the context of higher-level +middleware. Kokyu currently provides real-time scheduling and dispatching +services for TAO’s real-time CORBA Event Service, which mediates supplier-consumer +relationships between application operations. Kokyu consists primarily +of two cooperating infrastructure segments, illustrated in Figure 1: +<center> +<p><img SRC="kokyu1.jpg" height=285 width=489> +<br><b>Figure 1: Kokyu Scheduling and Dispatching Infrastructure</b></center> + +<ol> +<li> +A pluggable scheduling infrastructure with efficient support for adaptive +execution of diverse static, dynamic, and hybrid static/dynamic scheduling +heuristics.</li> + +<li> +A flexible dispatching infrastructure that allows composition of primitive +operating system and middleware mechanisms to enforce arbitrary scheduling +heuristics.</li> +</ol> +The scheduler is responsible for specifying how operation dispatch requests +are ordered, by assigning priority levels and rates to tasks, and producing +a configuration specification for the dispatching mechanism. The dispatcher +is responsible for enforcing the ordering of operation dispatches using +different threads, requests queues, and timers configured according to +the scheduler’s specification. The combined framework provides an implicit +projection of scheduling heuristics into appropriate dispatching infrastructure +configurations, so that the scheduling and dispatching infrastructure segments +can be optimized both separately and in combination. +<h3> +<a NAME="SchedFramework"></a>Strategized Scheduling framework</h3> +The Kokyu scheduling framework is designed to support a variety of scheduling +heuristics including RMS, EDF, MLF, and MUF. In addition, this framework +provides a common environment to compare systematically both existing and +new scheduling strategies. This flexibility is achieved in the Kokyu framework +via the Strategy pattern, which allows parts of the sequence of steps in +an algorithm to be replaced, thus providing interchangeable variations +within a consistent algorithmic framework. The Kokyu scheduling framework +uses the Strategy pattern to encapsulate a family of scheduling algorithms +within a fixed CORBA IDL interface, thereby enabling different strategies +to be configured independently from applications that use them. +<h3> +<a NAME="FlexDispatch"></a>Flexible Dispatching Framework</h3> +The right side of Figure 1 shows the essential features of Kokyu’s flexible +task dispatching infrastructure. Key features of the dispatching infrastructure +that are essential to performing our optimizations are as follows: +<p><b>Dispatching queues:</b> Each task is assigned by our strategized +Kokyu scheduling framework to a specific dispatching queue, each +of which has an associated queue number, a queueing discipline, and a unique +operating-system-specific priority for its single associated dispatching +thread. +<p><b>Dispatching threads:</b> Operating-system thread priorities decrease +as the queue number increases, so that the 0th queue is served by the highest +priority thread. Each dispatching thread removes the task from the head +of its queue and runs its entry point function to completion before retrieving +the next task to dispatch. Adapters can be applied to operations to intercept +and possibly short-circuit the entry-point upcall. In general, however, +the outermost operation entry point must complete on each dispatch. +<p><b>Queueing disciplines: </b>Dispatching thread priorities determine +which queue is active at any given time: the highest priority queue with +a task to dispatch is always active, preempting tasks in lower priority +queues. In addition, each queue may have a distinct discipline for determining +which of its enqueued tasks has the highest eligibility, and must ensure +the highest is at the head of the queue at the point when one is to be +dequeued. We consider three disciplines: +<ul> +<li> +Static – Tasks are ordered by a static subpriority value – results in FIFO +ordering if all static subpriorities are made the same; static queues at +different priority levels can be used to implement an RMS scheduling strategy.</li> + +<li> +Deadline – Tasks are ordered by time to deadline; a single deadline queue +can be used to implement the earliest deadline first (EDF) scheduling strategy.</li> + +<li> +Laxity – Tasks are ordered by slack time, or laxity – the time to deadline +minus the execution time; a single laxity queue can be used to implement +the minimum laxity first (MLF) scheduling strategy; laxity queues at different +priority levels can be used to implement the maximum urgency first (MUF) +scheduling strategy.</li> +</ul> +Any discipline for which a maximal eligibility may be selected can be employed +to manage a given dispatching queue in this approach. Scheduling strategies +can be constructed from one or more queues of each discipline alone, or +combinations of queues with different disciplines can be used. Figure 2 +illustrates the general queueing mechanism used by the dispatching modules +in the Kokyu dispatching framework. +<center> +<p><img SRC="kokyu2.jpg" height=176 width=779> +<p><b>Figure 2: Example Queueing Mechanism in a Kokyu Dispatching Module</b></center> + +<p>In addition, this figure shows how the output information provided by +the Kokyu scheduling framework is used to configure and operate a dispatching +module. During system initialization, each dispatching module obtains the +thread priority and dispatching type for each of its queues, typically +from the scheduling service’s output interface. Next, each queue is assigned +a unique dispatching priority number, a unique thread priority, and an +enumerated dispatching type. Finally, each dispatching module has an ordered +queue of pending dispatches per dispatching priority. To preserve QoS guarantees, +operations are inserted into the appropriate dispatching queue according +to their assigned dispatching priority. Operations within a dispatching +queue are ordered by their assigned dispatching subpriority. To minimize +priority inversions, operations are dispatched from the queue with the +highest thread priority, preempting any operation executing in a lower +priority thread. To minimize preemption overhead, there is no preemption +within a given priority queue. The following three values are defined for +the dispatching type: +<ul> +<li> +<b>STATIC DISPATCHING</b>: This type specifies a queue that only considers +the static portion of an operation’s dispatching subpriority.</li> + +<li> +<b>DEADLINE DISPATCHING</b>: This type specifies a queue that considers +the dynamic and static portions of an operation’s dispatching subpriority, +and updates the dynamic portion according to the time remaining until the +operation’s deadline.</li> + +<li> +<b>LAXITY DISPATCHING</b>: This type specifies a queue that considers the +dynamic and static portions of an operation’s dispatching subpriority, +and updates the dynamic portion according to the operation’s laxity.</li> +</ul> + +<h3> +<a NAME="KokyuEC"></a>Use of Kokyu within the TAO Real-time Event Channel(RTEC)</h3> +Figure 3 shows the sequence of operations that take place in the Kokyu +based dispatching module in the TAO RTEC. The client application registers +all relevant operations with the scheduler along with their real-time requirements. +This is done through the concept of an <font face="Courier New,Courier">RT_Info +</font>(see +TAO/orbsvcs/orbsvcs/RtecScheduler.idl) structure which is a structure that +contains the execution time, criticality, period, etc of an operation. +The client then calls <font face="Courier New,Courier">compute_schedule</font> +method on the scheduler. The scheduler creates a dependency graphs of all +operations and partitions operations into equivalence classes based on +the scheduling parameters supplied. The scheduler can be configured to +have any scheduling policy which determines the equivalence class partitioning +(queues) and possibly a partial ordering of operations within an equivalence +class (ordering within a queue). Once this is done, the scheduler has the +configuration information for the Kokyu dispatcher like the number of dispatch +queues, priorities for the threads processing each queue, etc. +<p>When the client calls <font face="Courier New,Courier">activate</font> +on the event channel, the EC inturn activates the Kokyu based EC dispatching +module. The EC dispatching module queries the dispatch configuration from +the scheduler and uses that to create the Kokyu dispatcher with the appropriate +number of lanes and threads. When an event is pushed into the EC, the EC +pushes the event to the appropriate consumers, who are subscribed to that +event. For each consumer, the EC queries the scheduler for the RT_Info +of that consumer. It then hands over the event to the Kokyu based dispatching +module. The dispatching module then enqueues the event into the appropriate +queue for processing by the thread watching that queue. +<center> +<p><img SRC="KokyuEC.jpg" height=784 width=716> +<p><b>Figure 3: Kokyu based dispatching module within TAO RTEC</b></center> + +<h3> +<a NAME="ConfigKokyuEC"></a>Configuration of RTEC to use Kokyu dispatching</h3> +<b>Static configuration</b>: In the <b>svc.conf</b> file, make sure you +have the following configuration for Kokyu dispatching. You can combine +this with other -ECxxx options. +<p><font face="Courier New,Courier">static EC_Factory "-ECdispatching kokyu +SCHED_FIFO -ECscheduling kokyu -ECfiltering kokyu"</font> +<p>To run the threads in the real-time FIFO class, use SCHED_FIFO. You +could use SCHED_RR and SCHED_OTHER also. +<br>The default is SCHED_FIFO. +<p>In your program, call +<p><font face="Courier New,Courier">TAO_EC_Kokyu_Factory::init_svcs ();</font> +<p>to statically create the EC Kokyu dispatching and other Kokyu related +modules. +<p><b>Dynamic configuration</b>: In the <b>svc.conf</b> file, make sure +you have the following configuration for Kokyu dispatching. You can combine +this with other -ECxxx options. +<p><font face="Courier New,Courier">dynamic EC_Factory Service_Object * +TAO_RTKokyuEvent:_make_TAO_EC_Kokyu_Factory() "-ECdispatching kokyu -ECscheduling +kokyu -ECfiltering kokyu"</font> +<h3> +<a NAME="KokyuDSRTCORBA"></a>Use of Kokyu within the Dynamic Scheduling +Real-time CORBA (DSRTCORBA) schedulers</h3> +An initial implementation of mechanisms to support DSRTCORBA schedulers +have been released. DSRTCORBA uses the concept of distributed threads, +which traverse multiple end systems giving the application the illusion +of a single logical thread executing an end-to-end task. The distributed +thread carries with it the scheduling parameters like importance, deadline, +etc so that it can get scheduled by a local scheduler on each endsystem. +The Kokyu DSRT dispatching framework is used as an enforcing mechanism. +<p>The DSRT schedulers are available in the directory $TAO_ROOT/examples/Kokyu_dsrt_schedulers. +They use the Kokyu DSRT +<br>dispatching classes present in $ACE_ROOT/Kokyu. These act as wrappers/adapters +around the Kokyu DSRT dispatcher. The Kokyu DSRT dispatcher is responsible +for scheduling threads which ask the dispatcher to schedule themselves. +Currently there are two implementations for the Kokyu DSRT dispatcher. +One uses a condition-variable based approach for scheduling threads and +the other manipulates priorities of threads and relies on the OS scheduler +for dispatching the threads appropriately. +<h4> +CV-based approach:</h4> +In this approach, it is assumed that the threads "yield" on a regular basis +to the scheduler by calling <tt>update_scheduling_segment</tt>. Only one +thread is running at any point in time. All the other threads are blocked +on a condition variable. When the currently running thread yields, it will +cause the condition variable to be signalled. All the eligible threads +are stored in a scheduler queue (rbtree), the most eligible thread determined +by the scheduling discipline. This approach has the drawback that it requires +a cooperative threading model, where threads yield voluntarily on a regular +basis. The application threads are responsible for doing this voluntary +yielding. +<h4> +OS-based approach:</h4> +This approach relies on the OS scheduler to do the actual thread dispatching. +The Kokyu DSRT dispatcher manipulates the priorities of the threads. The +scheduler maintains a queue (rbtree) of threads. The scheduler also has +an executive thread, which runs at the maximum available priority. This +thread runs in a continuous loop until the dispatcher is shut down. The +executive thread is responsible for selecting the most eligible thread +from the scheduler queue and bump up its priority if necessary while bumping +down the priority of the currently running thread, if it is not the most +eligible. There are four priority levels required for this mechanism to +work, listed in descending order of priorities. For example, a thread running +at <i>Active</i> priority will preempt a +<br>thread running at <i>Inactive</i> priority level. +<ol> +<li> +Executive priority - priority at which the scheduler executive thread runs.</li> + +<li> +Blocked priority - this is the priority to which threads about to block +on remote calls will be bumped up to.</li> + +<li> +Active priority - this is the priority to which the most eligible thread +is set to.</li> + +<li> +Inactive priority - this is the priority to which all threads except the +most eligible thread is set to.</li> +</ol> +As soon as a thread asks to be scheduled, a wrapper object is created and +inserted into the queue. This object carries the qos (sched params) associated +with that thread. A condition variable is signalled to inform the executive +thread that the queue is "dirty". The scheduler thread picks up the most +eligble one and sets its priority to <i>active</i> and sets the currently +running thread priority to +<br><i>inactive</i>. +<p>The drawback to this approach is that it relies on the OS scheduler +to dispatch the threads. Also, with the current implementation, there is +only one thread running at active priority and others are all at <i>inactive</i> +level. This will create undesirable effects with multi-processor systems, +which could select any one of the <i>inactive</i> level threads and this +could cause priority inversions. +<h3> +<a NAME="newDSRTSched"></a>How to write a new DSRT scheduler using Kokyu</h3> +One can use one of the schedulers as a starting point. The variation points +are +<ol> +<li> +The scheduler parameters that need to be propagated along with the service +context.</li> + +<li> +The QoS comparison function, that determines which thread is more eligible.</li> +</ol> +To aid (1), we have created a Svc_Ctxt_DSRT_QoS idl interface (see ./Kokyu_qos.pidl). +This interface currently has the necessary things to be propagated for +FP, MIF and MUF schedulers. This can be altered if necessary to accomodate +new sched params. The idea here is to let the IDL compiler generate the +marshalling code (including Any operators) so that these parameters can +be shipped across in the service context in an encapsulated CDR. +<p>To create customized QoS comparator functions, we used the idea of C++ +traits to let the user define customized comparator functions. For example, +the MIF scheduler uses the following traits class. +<p><tt> struct MIF_Scheduler_Traits</tt> +<br><tt> {</tt> +<br><tt> typedef RTScheduling::Current::IdType Guid_t;</tt> +<p><tt> struct _QoSDescriptor_t</tt> +<br><tt> {</tt> +<br><tt> typedef long Importance_t;</tt> +<br><tt> Importance_t importance_;</tt> +<br><tt> };</tt> +<p><tt> typedef _QoSDescriptor_t QoSDescriptor_t;</tt> +<p><tt> typedef Kokyu::MIF_Comparator<QoSDescriptor_t> +QoSComparator_t;</tt> +<p><tt> class _Guid_Hash</tt> +<br><tt> {</tt> +<br><tt> public:</tt> +<br><tt> u_long operator () (const Guid_t& +id)</tt> +<br><tt> {</tt> +<br><tt> return ACE::hash_pjw +((const char *) id.get_buffer (),</tt> +<br><tt> +id.length ());</tt> +<br><tt> }</tt> +<br><tt> };</tt> +<p><tt> typedef _Guid_Hash Guid_Hash;</tt> +<br><tt> };</tt> +<p>The idea of traits makes the Kokyu dispatcher more flexible in terms +of creating new schedulers. For example, the Kokyu classes do not care +about what concrete type Guid is. It could be an OctetSequence for some +applications, whereas it could be an int for some others. The exact type +is defined by the application (in this case, the MIF scheduler) using the +traits class. In the above traits class the Guid's type is defined to be +an octet sequence (indirectly). The Kokyu dispatcher expects the following +typedef's to +<br>be present in the traits class: +<p><tt>Guid_t - </tt>Type of GUID. +<br><tt>QoSDescriptor_t - </tt>aggregate for scheduler parameters +<br><tt>QoSComparator_t - </tt>used by the scheduler queue to determine +most eligible item +<br><tt>Guid_Hash - </tt>used by the internal hash map in the scheduler +to hash the guid. +<p>It is also expected that the following operator be defined for comparing +QoS parameters. This comparator function will be used by the scheduler +queue to determine the most eligible item in the queue. +<p><tt>QoSComparator_t::operator ()(const QoSDescriptor_t& qos1,</tt> +<br><tt> +const QoSDescriptor_t& qos2)</tt> +<h3> +<a NAME="DSRTCORBAvsRTEC"></a>Kokyu DSRTCORBA vs Kokyu RTEC</h3> +Currently we have separate interfaces for DSRTCORBA and RTEC dispatching +mechanisms. Once we get more use cases and experience, there is a possibility +of these getting merged in the future. The RTEC related dispatching interface +is in <tt>Kokyu::Dispatcher (Kokyu.h)</tt> and DSRTCORBA related dispatching +interface is in <tt>Kokyu::DSRT_Dispatcher (Kokyu_dsrt.h)</tt> +<h3> +<a NAME="Status"></a>Current status</h3> +Kokyu dispatching framework is available as a separate module under <tt><font size=+1>ACE_wrappers/Kokyu</font></tt> +as part of the <a href="http://deuce.doc.wustl.edu/Download.html">ACE/TAO +distribution</a>. Note that this module is not dependent on TAO, though +it is built on top of ACE. The TAO Event Channel uses the Strategy and +Service Configurator patterns to use configurable dispatching modules. +A Kokyu based EC dispatching module is available in the <tt><font size=+1>TAO/orbsvcs/orbsvcs/RTKokyuEvent</font></tt> +module. This module acts as an adapter between the Kokyu dispatcher and +the RTEC. +<p>Kokyu scheduling framework is available under the TAO source tree (<tt><font size=+1>TAO/orbsvcs/orbsvcs/Sched</font></tt>). +<p>An example using the RTEC Kokyu dispatching module is available under +<tt><font size=+1>TAO/orbsvcs/examples/RtEC/Kokyu</font></tt>. +<h3> +<a NAME="Future"></a>Future work</h3> + +<ol> +<li> +Currently there is no support for timers in the Kokyu dispatching module. +We plan to do this in the near future.</li> + +<li> +It looks like there is a general structure to the different schedulers. +May be this can be abstracted using templates or some similar mechanism.</li> + +<li> +Thread sched policy and sched scope are currently being passed explicitly +from the application to the scheduler. This can be changed later to get +this information from the ORB. This requires the usage of RTORB and the +actual values can be set using svc.conf parameters for RT_ORB_Loader.</li> + +<br> +<li> +See whether the approaches could be extended to multiprocessor systems.</li> +</ol> + +<h3> +<a NAME="Papers"></a>Papers on Kokyu</h3> + +<ol> +<li> +Christopher D. Gill, <a href="http://www.cse.wustl.edu/~cdgill/PDF/cdgill_dissertation.pdf">Dissertation:Flexible +Scheduling in Middleware for Distributed Rate-Based Real-Time Applications</a></li> + +<li> +Christopher D. Gill, David L. Levine, and Douglas C. Schmidt <a href="http://www.cse.wustl.edu/~cdgill/PDF/dynamic.pdf">The +Design and Performance of a Real-Time CORBA Scheduling Service</a>, Real-Time +Systems: the International Journal of Time-Critical Computing Systems, +special issue on Real-Time Middleware, guest editor Wei Zhao, March 2001, +Vol. 20 No. 2</li> + +<li> +Christopher D. Gill, Douglas C. Schmidt, and Ron Cytron, <a href="http://www.cs.wustl.edu/~schmidt/PDF/embedded_sched.pdf">Multi-Paradigm +Scheduling for Distributed Real-Time Embedded Computing</a>, IEEE Proceedings +Special Issue on Modeling and Design of Embedded Systems, Volume 91, Number +1, January 2003.</li> +</ol> + +</body> +</html> diff --git a/Kokyu/KokyuEC.jpg b/Kokyu/KokyuEC.jpg Binary files differnew file mode 100644 index 00000000000..8be16ef04c3 --- /dev/null +++ b/Kokyu/KokyuEC.jpg diff --git a/Kokyu/kokyu1.jpg b/Kokyu/kokyu1.jpg Binary files differnew file mode 100644 index 00000000000..094f5535d02 --- /dev/null +++ b/Kokyu/kokyu1.jpg diff --git a/Kokyu/kokyu2.jpg b/Kokyu/kokyu2.jpg Binary files differnew file mode 100644 index 00000000000..2c649bfc754 --- /dev/null +++ b/Kokyu/kokyu2.jpg |