lamellar 0.8.0

Lamellar is an asynchronous tasking runtime for HPC systems developed in RUST.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189

use crate::active_messaging::SyncSend;
use crate::array::iterator::consumer::*;
use crate::array::r#unsafe::private::UnsafeArrayInner;
// use crate::LamellarEnv;
// use crate::memregion::Dist;

use parking_lot::Mutex;
use std::collections::VecDeque;
use std::sync::atomic::AtomicUsize;
use std::sync::Arc;

impl UnsafeArrayInner {
    pub(crate) fn sched_static<C, AmO, O, I>(&self, cons: C) -> C::Handle
    where
        C: IterConsumer<AmOutput = AmO, Output = O, Item = I>,
        AmO: SyncSend + 'static,
        O: SyncSend + 'static,
        // I: SyncSend + 'static,
    {
        let mut reqs = VecDeque::new();
        if let Ok(_my_pe) = self.data.team.team_pe_id() {
            let num_workers = self.data.team.num_threads();
            let num_elems_local = cons.max_elems(self.num_elems_local());
            let elems_per_thread = 1.0f64.max(num_elems_local as f64 / num_workers as f64);
            // println!("static num_workers {:?} num_elems_local {:?} elems_per_thread {:?}", num_workers, num_elems_local, elems_per_thread);
            let mut worker = 0;
            while ((worker as f64 * elems_per_thread).round() as usize) < num_elems_local {
                let start_i = (worker as f64 * elems_per_thread).round() as usize;
                let end_i = ((worker + 1) as f64 * elems_per_thread).round() as usize;
                reqs.push_back(
                    self.data.task_group.exec_arc_am_local_inner(
                        cons.into_am(IterSchedule::Static(start_i, end_i)),
                    ),
                );

                worker += 1;
            }
        }
        cons.create_handle(self.data.inner().darc_rt_team().clone(), reqs)
    }

    pub(crate) fn sched_dynamic<C, AmO, O, I>(&self, cons: C) -> C::Handle
    where
        C: IterConsumer<AmOutput = AmO, Output = O, Item = I>,
        AmO: SyncSend + 'static,
        O: SyncSend + 'static,
        // I: SyncSend + 'static,
    {
        let mut reqs = VecDeque::new();
        if let Ok(_my_pe) = self.data.team.team_pe_id() {
            let num_workers = self.data.team.num_threads();
            let num_elems_local = cons.max_elems(self.num_elems_local());
            // println!("dynamic num_workers {:?} num_elems_local {:?}", num_workers, num_elems_local);

            let cur_i = Arc::new(AtomicUsize::new(0));
            // println!("ranges {:?}", ranges);
            for _ in 0..std::cmp::min(num_workers, num_elems_local) {
                reqs.push_back(self.data.task_group.exec_arc_am_local_inner(
                    cons.into_am(IterSchedule::Dynamic(cur_i.clone(), num_elems_local)),
                ));
            }
        }
        cons.create_handle(self.data.team.clone(), reqs)
    }

    pub(crate) fn sched_work_stealing<C, AmO, O, I>(&self, cons: C) -> C::Handle
    where
        C: IterConsumer<AmOutput = AmO, Output = O, Item = I>,
        AmO: SyncSend + 'static,
        O: SyncSend + 'static,
        // I: SyncSend + 'static,
    {
        let mut reqs = VecDeque::new();
        if let Ok(_my_pe) = self.data.team.team_pe_id() {
            let num_workers = self.data.team.num_threads();
            let num_elems_local = cons.max_elems(self.num_elems_local());
            let elems_per_thread = 1.0f64.max(num_elems_local as f64 / num_workers as f64);
            // println!("work stealing num_workers {:?} num_elems_local {:?} elems_per_thread {:?}", num_workers, num_elems_local, elems_per_thread);
            let mut worker = 0;
            let mut siblings = Vec::new();
            while ((worker as f64 * elems_per_thread).round() as usize) < num_elems_local {
                let start_i = (worker as f64 * elems_per_thread).round() as usize;
                let end_i = ((worker + 1) as f64 * elems_per_thread).round() as usize;
                siblings.push(IterWorkStealer {
                    range: Arc::new(Mutex::new((start_i, end_i))),
                });
                worker += 1;
            }
            for sibling in &siblings {
                reqs.push_back(self.data.task_group.exec_arc_am_local_inner(cons.into_am(
                    IterSchedule::WorkStealing(sibling.clone(), siblings.clone()),
                )))
            }
        }
        cons.create_handle(self.data.team.clone(), reqs)
    }

    pub(crate) fn sched_guided<C, AmO, O, I>(&self, cons: C) -> C::Handle
    where
        C: IterConsumer<AmOutput = AmO, Output = O, Item = I>,
        AmO: SyncSend + 'static,
        O: SyncSend + 'static,
        // I: SyncSend + 'static,
    {
        let mut reqs = VecDeque::new();
        if let Ok(_my_pe) = self.data.team.team_pe_id() {
            let num_workers = self.data.team.num_threads();
            let num_elems_local_orig = cons.max_elems(self.num_elems_local());
            let mut num_elems_local = num_elems_local_orig as f64;
            let mut elems_per_thread = num_elems_local / num_workers as f64;
            // println!("guided num_workers {:?} num_elems_local_orig {:?} num_elems_local {:?} elems_per_thread {:?}", num_workers, num_elems_local_orig, num_elems_local, elems_per_thread);
            let mut ranges = Vec::new();
            let mut cur_i = 0;
            let mut i;
            while elems_per_thread > 100.0 && cur_i < num_elems_local_orig {
                num_elems_local = num_elems_local / 1.61; //golden ratio
                let start_i = cur_i;
                let end_i = std::cmp::min(
                    cur_i + num_elems_local.round() as usize,
                    num_elems_local_orig,
                );
                i = 0;
                while cur_i < end_i {
                    ranges.push((
                        start_i + (i as f64 * elems_per_thread).round() as usize,
                        start_i + ((i + 1) as f64 * elems_per_thread).round() as usize,
                    ));
                    i += 1;
                    cur_i = start_i + (i as f64 * elems_per_thread).round() as usize;
                }
                elems_per_thread = num_elems_local / num_workers as f64;
            }
            if elems_per_thread < 1.0 {
                elems_per_thread = 1.0;
            }
            i = 0;
            let start_i = cur_i;
            while cur_i < num_elems_local_orig {
                ranges.push((
                    start_i + (i as f64 * elems_per_thread).round() as usize,
                    start_i + ((i + 1) as f64 * elems_per_thread).round() as usize,
                ));
                i += 1;
                cur_i = start_i + (i as f64 * elems_per_thread).round() as usize;
            }
            let range_i = Arc::new(AtomicUsize::new(0));
            // println!("ranges {:?}", ranges);
            for _ in 0..std::cmp::min(num_workers, num_elems_local_orig) {
                reqs.push_back(self.data.task_group.exec_arc_am_local_inner(
                    cons.into_am(IterSchedule::Chunk(ranges.clone(), range_i.clone())),
                ));
            }
        }
        cons.create_handle(self.data.team.clone(), reqs)
    }

    pub(crate) fn sched_chunk<C, AmO, O, I>(&self, cons: C, chunk_size: usize) -> C::Handle
    where
        C: IterConsumer<AmOutput = AmO, Output = O, Item = I>,
        AmO: SyncSend + 'static,
        O: SyncSend + 'static,
        // I: SyncSend + 'static,
    {
        let mut reqs = VecDeque::new();
        if let Ok(_my_pe) = self.data.team.team_pe_id() {
            let num_workers = self.data.team.num_threads();
            let num_elems_local = cons.max_elems(self.num_elems_local());
            let mut ranges = Vec::new();
            let mut cur_i = 0;
            let mut num_chunks = 0;
            while cur_i < num_elems_local {
                ranges.push((cur_i, cur_i + chunk_size));
                cur_i += chunk_size;
                num_chunks += 1;
            }

            // println!("chunk num_workers {:?} num_elems_local {:?} num_chunks {:?}", num_workers, num_elems_local, num_chunks);

            let range_i = Arc::new(AtomicUsize::new(0));
            // println!("ranges {:?}", ranges);
            for _ in 0..std::cmp::min(num_workers, num_chunks) {
                reqs.push_back(self.data.task_group.exec_arc_am_local_inner(
                    cons.into_am(IterSchedule::Chunk(ranges.clone(), range_i.clone())),
                ));
            }
        }
        cons.create_handle(self.data.team.clone(), reqs)
    }
}