1#!/usr/bin/python3
2# Copyright 2017 Facebook.
3# Licensed under the same terms as memcached itself.
4
5import argparse
6import socket
7import sys
8import re
9import traceback
10from time import sleep, time
11
12parser = argparse.ArgumentParser(description="daemon for rebalancing slabs")
13parser.add_argument("--host", help="host to connect to",
14        default="localhost:11211", metavar="HOST:PORT")
15parser.add_argument("-s", "--sleep", help="seconds between runs",
16                    type=int, default="1")
17parser.add_argument("-v", "--verbose", action="store_true")
18parser.add_argument("-a", "--automove", action="store_true", default=False,
19                    help="enable automatic page rebalancing")
20parser.add_argument("-w", "--window", type=int, default="30",
21                    help="rolling window size for decision history")
22parser.add_argument("-r", "--ratio", type=float, default=0.8,
23                    help="ratio limiting distance between low/high class ages")
24parser.add_argument("-f", "--free", type=float, default=0.005,
25                    help="free chunks/pages buffer ratio")
26parser.add_argument("-z", "--size", type=int, default=512,
27                    help="item size cutoff for storage")
28
29args = parser.parse_args()
30
31host, port = args.host.split(':')
32
33MIN_PAGES_FOR_SOURCE = 2
34MIN_PAGES_FOR_RECLAIM = 2.5
35MIN_PAGES_FREE = 1.5
36MEMCHECK_PERIOD = 60
37
38def window_check(history, sid, key):
39    total = 0
40    for window in history['w']:
41        s = window.get(sid)
42        if s and s.get(key):
43            total += s.get(key)
44    return total
45
46def window_key_check(history, key):
47    total = 0
48    for window in history['w']:
49        v = window.get(key)
50        if v:
51            total += v
52    return total
53
54
55def determine_move(history, stats, diffs, memfree):
56    """ Figure out of a page move is in order.
57
58    - Use as much memory as possible to hold items, reducing the load on
59      flash.
60    - tries to keep the global free page pool inbetween poolmin/poolmax.
61    - avoids flapping as much as possible:
62      - only pull pages off of a class if it hasn't recently evicted or allocated pages.
63      - only pull pages off if a sufficient number of free chunks are available.
64      - if global pool is below minimum remove pages from oldest large class.
65      - if global pool is above maximum, move pages to youngest large class.
66    - extstore manages a desired number of free chunks in each slab class.
67    - automover adjusts above limits once per minute based on current sizes.
68    - if youngest is below the age ratio limit of oldest, move a page to it.
69    """
70    # rotate windows
71    history['w'].append({})
72    if (len(history['w']) > args.window):
73        history['w'].pop(0)
74    w = history['w'][-1]
75    oldest = (-1, 0)
76    youngest = (-1, sys.maxsize)
77    too_free = False
78    # Most bytes free
79    decision = (-1, -1)
80    if int(stats['slab_global_page_pool']) < memfree[0] / 2:
81        w['slab_pool_low'] = 1
82    if int(stats['slab_global_page_pool']) > memfree[0]:
83        w['slab_pool_high'] = 1
84    if args.verbose:
85        print("global pool: [{}]".format(stats['slab_global_page_pool']))
86
87    pool_low = window_key_check(history, 'slab_pool_low')
88    for sid, slab in diffs.items():
89        small_slab = False
90        free_enough = False
91        # Only balance larger slab classes
92        if slab['chunk_size'] < args.size:
93            small_slab = True
94
95        w[sid] = {}
96        if 'evicted_d' not in slab or 'total_pages_d' not in slab:
97            continue
98        # mark this window as dirty if total pages increases or evictions
99        # happened
100        if slab['total_pages_d'] > 0:
101            w[sid]['dirty'] = 1
102        if slab['evicted_d'] > 0:
103            w[sid]['dirty'] = 1
104            w[sid]['ev'] = 1
105        if slab['free_chunks'] > memfree[sid]:
106            free_enough = True
107        if memfree[sid] > 0 and slab['free_chunks'] > (memfree[sid] * 2):
108            w[sid]['excess_free'] = 1
109        w[sid]['age'] = slab['age']
110        age = window_check(history, sid, 'age') / len(history['w'])
111
112        # if > 2.5 pages free, and not dirty, reassign to global page pool
113        if slab['free_chunks'] > slab['chunks_per_page'] * MIN_PAGES_FOR_RECLAIM and too_free == False:
114            dirt = window_check(history, sid, 'dirty')
115            excess = window_check(history, sid, 'excess_free')
116            if small_slab == True and dirt == 0:
117                # If we're a small slab, don't hang on to free memory forever.
118                decision = (sid, 0)
119                too_free = True
120            elif small_slab == False and dirt == 0 \
121                    and excess >= len(history['w']):
122                decision = (sid, 0)
123                too_free = True
124
125        # are we the oldest slab class? (and a valid target)
126        # don't consider for young if we've recently given it unused memory
127        if small_slab == False:
128            if age > oldest[1] and slab['total_pages'] > MIN_PAGES_FOR_SOURCE:
129                oldest = (sid, age)
130            if age < youngest[1] and slab['total_pages'] > 0 \
131                    and window_check(history, sid, 'excess_free') < len(history['w']) \
132                    and not (window_check(history, sid, 'relaxed') and free_enough):
133                youngest = (sid, age)
134
135
136    if w.get('slab_pool_high') and youngest[0] != -1:
137        # if global pool is too high, feed youngest large class.
138        if slab['free_chunks'] <= memfree[youngest[0]]:
139            decision = (0, youngest[0])
140        w[youngest[0]]['relaxed'] = 1
141    elif too_free == False and pool_low and oldest[0] != -1:
142        # if pool is too low, take from oldest large class.
143        if args.verbose:
144            print("oldest:  [class: {}] [age: {:.2f}]".format(int(oldest[0]), oldest[1]))
145        decision = (oldest[0], 0)
146    elif too_free == False and youngest[0] != -1 and oldest[0] != -1 and youngest[0] != oldest[0]:
147        # youngest is outside of the tolerance ratio, move a page around.
148        if args.verbose:
149            print("old:   [class: {}] [age: {:.2f}]\nyoung: [class: {}] [age: {:.2f}]".format(
150                int(oldest[0]), oldest[1], int(youngest[0]), youngest[1]))
151
152        slab = diffs[youngest[0]]
153        #print("F:{} L:{} Y:{} R:{}".format(slab['free_chunks'], memfree[youngest[0]], int(youngest[1]), int(oldest[1] * args.ratio)))
154        if youngest[1] < oldest[1] * args.ratio:
155            w[youngest[0]]['relaxed'] = 1
156            if slab['free_chunks'] <= memfree[youngest[0]]:
157                decision = (0, youngest[0])
158
159    if (len(history['w']) >= args.window):
160        return decision
161    return (-1, -1)
162
163
164def run_move(s, decision):
165    s.write("slabs reassign " + str(decision[0]) + " " + str(decision[1]) + "\r\n")
166    line = s.readline().rstrip()
167    if args.verbose:
168        print("move result:", line)
169
170
171def diff_stats(before, after):
172    """ fills out "diffs" as deltas between before/after,
173    and "totals" as the sum of all slab classes.
174    "_d" postfix to keys means the delta between before/after.
175    non-postfix keys are total as of 'after's reading.
176    """
177    diffs = {}
178    totals = {}
179    for slabid in after.keys():
180        sb = before.get(slabid)
181        sa = after.get(slabid)
182        if not (sb and sa):
183            continue
184        slab = sa.copy()
185        for k in sa.keys():
186            if k not in sb:
187                continue
188            if k not in totals:
189                totals[k] = 0
190                totals[k + '_d'] = 0
191            if k + '_d' not in slab:
192                slab[k + '_d'] = 0
193            if re.search(r"^\d+$", sa[k]):
194                totals[k] += int(sa[k])
195                slab[k] = int(sa[k])
196                slab[k + '_d'] = int(sa[k]) - int(sb[k])
197                totals[k + '_d'] += int(sa[k]) - int(sb[k])
198        slab['slab'] = slabid
199        diffs[slabid] = slab
200    return (diffs, totals)
201
202
203def read_slab_stats(s):
204    slabs = {}
205    for statcmd in ['items', 'slabs']:
206        #print("stat cmd: " + statcmd)
207        # FIXME: Formatting
208        s.write("stats " + statcmd + "\r\n")
209        while True:
210            line = s.readline().rstrip()
211            if line.startswith("END"):
212                break
213
214            m = re.match(r"^STAT (?:items:)?(\d+):(\S+) (\S+)", line)
215            if m:
216                (slab, var, val) = m.groups()
217                if slab not in slabs:
218                    slabs[slab] = {}
219                slabs[slab][var] = val
220            #print("line: " + line)
221    return slabs
222
223
224# HACK: lets look at 'evictions' being nonzero to indicate memory filled at some point.
225def read_stats(s):
226    stats = {}
227    s.write("stats\r\n")
228    while True:
229        line = s.readline().rstrip()
230        if line.startswith("END"):
231            break
232
233        m = re.match(r"^STAT (\S+) (\S+)", line)
234        if m:
235            (key, val) = m.groups()
236            stats[key] = val
237    return stats
238
239
240def pct(num, divisor):
241    if not divisor:
242        return 0
243    return (num / divisor)
244
245
246def show_detail(diffs, totals):
247    """ just a pretty printer for some extra data """
248    print("\n  {:2s}: {:8s} (pct  ) {:10s} (pct    ) {:6s} (pct)   {:6s}".format('sb',
249                'evicted', 'items', 'pages', 'age'))
250
251    for sid, slab in diffs.items():
252        if 'evicted_d' not in slab:
253            continue
254        print("  {:2d}: {:8d} ({:.2f}%) {:10d} ({:.4f}%) {:6d} ({:.2f}%) {:6d}".format(
255              int(sid), slab['evicted_d'], pct(slab['evicted_d'], totals['evicted_d']),
256              slab['number'], pct(slab['number'], totals['number']),
257              slab['total_pages'], pct(slab['total_pages'],
258              totals['total_pages']),
259              slab['age']))
260
261def memfree_check(s, diffs, totals):
262    info = {}
263    # manage about this many free chunks in each slab class.
264    for sid, slab in diffs.items():
265        if sid == 0:
266            continue
267        hold_free = int((slab['used_chunks'] + slab['free_chunks']) * args.free)
268        # Hold a minimum of 1.5 pages so page moves are unlikely to lose items.
269        if slab['chunks_per_page'] * MIN_PAGES_FREE > hold_free:
270            hold_free = int(slab['chunks_per_page'] * MIN_PAGES_FREE)
271        info[sid] = hold_free
272        # TODO: only adjust if different?
273        s.write("extstore free_memchunks {} {}\r\n".format(sid, hold_free))
274        s.readline()
275
276    # how many pages to leave in the global pool.
277    info[0] = int(totals['total_pages'] * args.free)
278    return info
279
280
281stats_pre = {}
282history = { 'w': [{}] }
283memfree = { 0: 2 }
284last_memfree_check = 0
285while True:
286    try:
287        with socket.create_connection((host, port), 5) as c:
288            s = c.makefile(mode="rw", buffering=1)
289            s.write("slabs automove 0\r\n")
290            print(s.readline().rstrip())
291            while True:
292                stats_post = read_slab_stats(s)
293                stats = read_stats(s)
294                (diffs, totals) = diff_stats(stats_pre, stats_post)
295                #if args.verbose:
296                #    show_detail(diffs, totals)
297                if int(stats['evictions']) > 0:
298                    if (last_memfree_check < time() - 60) and totals.get('total_pages'):
299                        memfree = memfree_check(s, diffs, totals)
300                        last_memfree_check = time()
301                    decision = (-1, -1)
302                    decision = determine_move(history, stats, diffs, memfree)
303                    if int(decision[0]) > 0 and int(decision[1]) >= 0:
304                        print("moving page from, to:", decision)
305                        if args.automove:
306                            run_move(s, decision)
307
308                # Minimize sleeping if we just moved a page to global pool.
309                # Improves responsiveness during flushes/quick changes.
310                if decision[1] == 0:
311                    continue
312                else:
313                    sleep(args.sleep)
314                stats_pre = stats_post
315    except:
316        err = sys.exc_info()
317        print("disconnected:", err[0], err[1])
318        traceback.print_exc()
319        stats_pre = {}
320        history = { 'w': [{}] }
321        sleep(args.sleep)
322
323