source: libcfa/src/concurrency/ready_queue.cfa

Last change on this file was 31c967b, checked in by Thierry Delisle <tdelisle@…>, 2 months ago

Changed ready-queue so I can easily change the averaging algorithm.
Changed averaging to use logscale.

  • Property mode set to 100644
File size: 8.4 KB
Line 
1//
2// Cforall Version 1.0.0 Copyright (C) 2019 University of Waterloo
3//
4// The contents of this file are covered under the licence agreement in the
5// file "LICENCE" distributed with Cforall.
6//
7// ready_queue.cfa --
8//
9// Author           : Thierry Delisle
10// Created On       : Mon Nov dd 16:29:18 2019
11// Last Modified By :
12// Last Modified On :
13// Update Count     :
14//
15
16#define __cforall_thread__
17#define _GNU_SOURCE
18
19// #define __CFA_DEBUG_PRINT_READY_QUEUE__
20
21
22#define USE_AWARE_STEALING
23
24#include "bits/defs.hfa"
25#include "device/cpu.hfa"
26#include "kernel/cluster.hfa"
27#include "kernel/private.hfa"
28
29// #include <errno.h>
30// #include <unistd.h>
31
32#include "ready_subqueue.hfa"
33
34static const size_t cache_line_size = 64;
35
36#if !defined(__CFA_NO_STATISTICS__)
37        #define __STATS(...) __VA_ARGS__
38#else
39        #define __STATS(...)
40#endif
41
42static inline struct thread$ * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats));
43static inline struct thread$ * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats));
44static inline struct thread$ * search(struct cluster * cltr);
45
46//=======================================================================
47// Cforall Ready Queue used for scheduling
48//=======================================================================
49// void ?{}(__ready_queue_t & this) with (this) {
50//      lanes.data   = 0p;
51//      lanes.tscs   = 0p;
52//      lanes.caches = 0p;
53//      lanes.count  = 0;
54// }
55
56// void ^?{}(__ready_queue_t & this) with (this) {
57//      free(lanes.data);
58//      free(lanes.tscs);
59//      free(lanes.caches);
60// }
61
62//-----------------------------------------------------------------------
63__attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->sched) {
64        struct processor * const proc = kernelTLS().this_processor;
65        const bool external = (!proc) || (cltr != proc->cltr);
66        const bool remote   = hint == UNPARK_REMOTE;
67        const size_t lanes_count = readyQ.count;
68
69        /* paranoid */ verify( __shard_factor.readyq > 0 );
70        /* paranoid */ verify( lanes_count > 0 );
71
72        unsigned i;
73        if( external || remote ) {
74                // Figure out where thread was last time and make sure it's valid
75                /* paranoid */ verify(thrd->preferred >= 0);
76                unsigned start = thrd->preferred * __shard_factor.readyq;
77                if(start < lanes_count) {
78                        do {
79                                unsigned r = __tls_rand();
80                                i = start + (r % __shard_factor.readyq);
81                                /* paranoid */ verify( i < lanes_count );
82                                // If we can't lock it retry
83                        } while( !__atomic_try_acquire( &readyQ.data[i].l.lock ) );
84                } else {
85                        do {
86                                i = __tls_rand() % lanes_count;
87                        } while( !__atomic_try_acquire( &readyQ.data[i].l.lock ) );
88                }
89        } else {
90                do {
91                        unsigned r = proc->rdq.its++;
92                        i = proc->rdq.id + (r % __shard_factor.readyq);
93                        /* paranoid */ verify( i < lanes_count );
94                        // If we can't lock it retry
95                } while( !__atomic_try_acquire( &readyQ.data[i].l.lock ) );
96        }
97
98        // Actually push it
99        push(readyQ.data[i], thrd);
100
101        // Unlock and return
102        __atomic_unlock( &readyQ.data[i].l.lock );
103
104        #if !defined(__CFA_NO_STATISTICS__)
105                if(unlikely(external || remote)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED);
106                else __tls_stats()->ready.push.local.success++;
107        #endif
108}
109
110__attribute__((hot)) struct thread$ * pop_fast(struct cluster * cltr) with (cltr->sched) {
111        const size_t lanes_count = readyQ.count;
112
113        /* paranoid */ verify( __shard_factor.readyq > 0 );
114        /* paranoid */ verify( lanes_count > 0 );
115        /* paranoid */ verify( kernelTLS().this_processor );
116        /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes_count );
117
118        struct processor * const proc = kernelTLS().this_processor;
119        unsigned this = proc->rdq.id;
120        /* paranoid */ verify( this < lanes_count );
121        __cfadbg_print_safe(ready_queue, "Kernel : pop from %u\n", this);
122
123        // Figure out the current cache is
124        const unsigned this_cache = cache_id(cltr, this / __shard_factor.readyq);
125        const unsigned long long ctsc = rdtscl();
126
127        if(proc->rdq.target == UINT_MAX) {
128                uint64_t chaos = __tls_rand();
129                unsigned ext = chaos & 0xff;
130                unsigned other  = (chaos >> 8) % (lanes_count);
131
132                if(ext < 3 || __atomic_load_n(&caches[other / __shard_factor.readyq].id, __ATOMIC_RELAXED) == this_cache) {
133                        proc->rdq.target = other;
134                }
135        }
136        else {
137                const unsigned target = proc->rdq.target;
138                __cfadbg_print_safe(ready_queue, "Kernel : %u considering helping %u, tcsc %llu\n", this, target, readyQ.tscs[target].t.tv);
139                /* paranoid */ verify( readyQ.tscs[target].t.tv != ULLONG_MAX );
140                if(target < lanes_count) {
141                        const __readyQ_avg_t cutoff = calc_cutoff(ctsc, proc->rdq.id, lanes_count, cltr->sched.readyQ.data, cltr->sched.readyQ.tscs, __shard_factor.readyq, true);
142                        const __readyQ_avg_t age = moving_average(ctsc, readyQ.tscs[target].t.tv, readyQ.tscs[target].t.ma, false);
143                        __cfadbg_print_safe(ready_queue, "Kernel : Help attempt on %u from %u, age %'llu vs cutoff %'llu, %s\n", target, this, age, cutoff, age > cutoff ? "yes" : "no");
144                        if(age > cutoff) {
145                                thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help));
146                                if(t) return t;
147                        }
148                }
149                proc->rdq.target = UINT_MAX;
150        }
151
152        for(__shard_factor.readyq) {
153                unsigned i = this + (proc->rdq.itr++ % __shard_factor.readyq);
154                if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
155        }
156
157        // All lanes where empty return 0p
158        return 0p;
159
160}
161__attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) {
162        unsigned i = __tls_rand() % (cltr->sched.readyQ.count);
163        return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal));
164}
165__attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) {
166        return search(cltr);
167}
168
169//=======================================================================
170// Various Ready Queue utilities
171//=======================================================================
172// these function work the same or almost the same
173// whether they are using work-stealing or relaxed fifo scheduling
174
175//-----------------------------------------------------------------------
176// try to pop from a lane given by index w
177static inline struct thread$ * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->sched) {
178        /* paranoid */ verify( w < readyQ.count );
179        __STATS( stats.attempt++; )
180
181        // Get relevant elements locally
182        __intrusive_lane_t & lane = readyQ.data[w];
183
184        // If list looks empty retry
185        if( is_empty(lane) ) {
186                return 0p;
187        }
188
189        // If we can't get the lock retry
190        if( !__atomic_try_acquire(&lane.l.lock) ) {
191                return 0p;
192        }
193
194        // If list is empty, unlock and retry
195        if( is_empty(lane) ) {
196                __atomic_unlock(&lane.l.lock);
197                return 0p;
198        }
199
200        // Actually pop the list
201        struct thread$ * thrd;
202        unsigned long long ts_prev = ts(lane);
203        unsigned long long ts_next;
204        [thrd, ts_next] = pop(lane);
205
206        /* paranoid */ verify(thrd);
207        /* paranoid */ verify(ts_next);
208        /* paranoid */ verify(lane.l.lock);
209
210        // Unlock and return
211        __atomic_unlock(&lane.l.lock);
212
213        // Update statistics
214        __STATS( stats.success++; )
215
216        touch_tsc(readyQ.tscs, w, ts_prev, ts_next, true);
217
218        thrd->preferred = w / __shard_factor.readyq;
219
220        // return the popped thread
221        return thrd;
222}
223
224//-----------------------------------------------------------------------
225// try to pop from any lanes making sure you don't miss any threads push
226// before the start of the function
227static inline struct thread$ * search(struct cluster * cltr) {
228        const size_t lanes_count = cltr->sched.readyQ.count;
229        /* paranoid */ verify( lanes_count > 0 );
230        unsigned count = __atomic_load_n( &lanes_count, __ATOMIC_RELAXED );
231        unsigned offset = __tls_rand();
232        for(i; count) {
233                unsigned idx = (offset + i) % count;
234                struct thread$ * thrd = try_pop(cltr, idx __STATS(, __tls_stats()->ready.pop.search));
235                if(thrd) {
236                        return thrd;
237                }
238        }
239
240        // All lanes where empty return 0p
241        return 0p;
242}
243
244//-----------------------------------------------------------------------
245// get preferred ready for new thread
246unsigned ready_queue_new_preferred() {
247        unsigned pref = UINT_MAX;
248        if(struct thread$ * thrd = publicTLS_get( this_thread )) {
249                pref = thrd->preferred;
250        }
251
252        return pref;
253}
254
255//-----------------------------------------------------------------------
256// Given 2 indexes, pick the list with the oldest push an try to pop from it
257static inline struct thread$ * try_pop(struct cluster * cltr, unsigned i, unsigned j __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->sched) {
258        // Pick the bet list
259        int w = i;
260        if( __builtin_expect(!is_empty(readyQ.data[j]), true) ) {
261                w = (ts(readyQ.data[i]) < ts(readyQ.data[j])) ? i : j;
262        }
263
264        return try_pop(cltr, w __STATS(, stats));
265}
Note: See TracBrowser for help on using the repository browser.