X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/72ed2e6e260d8e12e49716a261b6306d8de13e8d..d63d6ba645b8e34cf9d248bd0407fab0b9ef6534:/sdk/python/arvados/keep.py diff --git a/sdk/python/arvados/keep.py b/sdk/python/arvados/keep.py index 44e9157767..cbe96ffa2f 100644 --- a/sdk/python/arvados/keep.py +++ b/sdk/python/arvados/keep.py @@ -15,6 +15,7 @@ from builtins import object import collections import datetime import hashlib +import errno import io import logging import math @@ -26,8 +27,11 @@ import socket import ssl import sys import threading +import resource from . import timer import urllib.parse +import traceback +import weakref if sys.version_info >= (3, 0): from io import BytesIO @@ -39,6 +43,7 @@ import arvados.config as config import arvados.errors import arvados.retry as retry import arvados.util +import arvados.diskcache _logger = logging.getLogger('arvados.keep') global_client_object = None @@ -174,11 +179,63 @@ class Keep(object): return Keep.global_client_object().put(data, **kwargs) class KeepBlockCache(object): - # Default RAM cache is 256MiB - def __init__(self, cache_max=(256 * 1024 * 1024)): + def __init__(self, cache_max=0, max_slots=0, disk_cache=False, disk_cache_dir=None): self.cache_max = cache_max self._cache = [] self._cache_lock = threading.Lock() + self._max_slots = max_slots + self._disk_cache = disk_cache + self._disk_cache_dir = disk_cache_dir + self._cache_updating = threading.Condition(self._cache_lock) + + if self._disk_cache and self._disk_cache_dir is None: + self._disk_cache_dir = os.path.join(os.path.expanduser("~"), ".cache", "arvados", "keep") + os.makedirs(self._disk_cache_dir, mode=0o700, exist_ok=True) + + if self._max_slots == 0: + if self._disk_cache: + # Each block uses two file descriptors, one used to + # open it initially and hold the flock(), and a second + # hidden one used by mmap(). + # + # Set max slots to 1/8 of maximum file handles. This + # means we'll use at most 1/4 of total file handles. + # + # NOFILE typically defaults to 1024 on Linux so this + # is 128 slots (256 file handles), which means we can + # cache up to 8 GiB of 64 MiB blocks. This leaves + # 768 file handles for sockets and other stuff. + # + # When we want the ability to have more cache (e.g. in + # arv-mount) we'll increase rlimit before calling + # this. + self._max_slots = int(resource.getrlimit(resource.RLIMIT_NOFILE)[0] / 8) + else: + # RAM cache slots + self._max_slots = 512 + + if self.cache_max == 0: + if self._disk_cache: + fs = os.statvfs(self._disk_cache_dir) + # Calculation of available space incorporates existing cache usage + existing_usage = arvados.diskcache.DiskCacheSlot.cache_usage(self._disk_cache_dir) + avail = (fs.f_bavail * fs.f_bsize + existing_usage) / 4 + maxdisk = int((fs.f_blocks * fs.f_bsize) * 0.10) + # pick smallest of: + # 10% of total disk size + # 25% of available space + # max_slots * 64 MiB + self.cache_max = min(min(maxdisk, avail), (self._max_slots * 64 * 1024 * 1024)) + else: + # 256 MiB in RAM + self.cache_max = (256 * 1024 * 1024) + + self.cache_max = max(self.cache_max, 64 * 1024 * 1024) + + if self._disk_cache: + self._cache = arvados.diskcache.DiskCacheSlot.init_cache(self._disk_cache_dir, self._max_slots) + self.cap_cache() + class CacheSlot(object): __slots__ = ("locator", "ready", "content") @@ -202,19 +259,51 @@ class KeepBlockCache(object): else: return len(self.content) + def evict(self): + self.content = None + return self.gone() + + def gone(self): + return (self.content is None) + + def _resize_cache(self, cache_max, max_slots): + # Try and make sure the contents of the cache do not exceed + # the supplied maximums. + + # Select all slots except those where ready.is_set() and content is + # None (that means there was an error reading the block). + self._cache = [c for c in self._cache if not (c.ready.is_set() and c.content is None)] + sm = sum([slot.size() for slot in self._cache]) + while len(self._cache) > 0 and (sm > cache_max or len(self._cache) > max_slots): + for i in range(len(self._cache)-1, -1, -1): + # start from the back, find a slot that is a candidate to evict + if self._cache[i].ready.is_set(): + sz = self._cache[i].size() + + # If evict returns false it means the + # underlying disk cache couldn't lock the file + # for deletion because another process was using + # it. Don't count it as reducing the amount + # of data in the cache, find something else to + # throw out. + if self._cache[i].evict(): + sm -= sz + + # check to make sure the underlying data is gone + if self._cache[i].gone(): + # either way we forget about it. either the + # other process will delete it, or if we need + # it again and it is still there, we'll find + # it on disk. + del self._cache[i] + break + + def cap_cache(self): '''Cap the cache size to self.cache_max''' - with self._cache_lock: - # Select all slots except those where ready.is_set() and content is - # None (that means there was an error reading the block). - self._cache = [c for c in self._cache if not (c.ready.is_set() and c.content is None)] - sm = sum([slot.size() for slot in self._cache]) - while len(self._cache) > 0 and sm > self.cache_max: - for i in range(len(self._cache)-1, -1, -1): - if self._cache[i].ready.is_set(): - del self._cache[i] - break - sm = sum([slot.size() for slot in self._cache]) + with self._cache_updating: + self._resize_cache(self.cache_max, self._max_slots) + self._cache_updating.notify_all() def _get(self, locator): # Test if the locator is already in the cache @@ -226,6 +315,12 @@ class KeepBlockCache(object): del self._cache[i] self._cache.insert(0, n) return n + if self._disk_cache: + # see if it exists on disk + n = arvados.diskcache.DiskCacheSlot.get_from_disk(locator, self._disk_cache_dir) + if n is not None: + self._cache.insert(0, n) + return n return None def get(self, locator): @@ -235,16 +330,66 @@ class KeepBlockCache(object): def reserve_cache(self, locator): '''Reserve a cache slot for the specified locator, or return the existing slot.''' - with self._cache_lock: + with self._cache_updating: n = self._get(locator) if n: return n, False else: # Add a new cache slot for the locator - n = KeepBlockCache.CacheSlot(locator) + self._resize_cache(self.cache_max, self._max_slots-1) + while len(self._cache) >= self._max_slots: + # If there isn't a slot available, need to wait + # for something to happen that releases one of the + # cache slots. Idle for 200 ms or woken up by + # another thread + self._cache_updating.wait(timeout=0.2) + self._resize_cache(self.cache_max, self._max_slots-1) + + if self._disk_cache: + n = arvados.diskcache.DiskCacheSlot(locator, self._disk_cache_dir) + else: + n = KeepBlockCache.CacheSlot(locator) self._cache.insert(0, n) return n, True + def set(self, slot, blob): + try: + slot.set(blob) + return + except OSError as e: + if e.errno == errno.ENOMEM: + # Reduce max slots to current - 4, cap cache and retry + with self._cache_lock: + self._max_slots = max(4, len(self._cache) - 4) + elif e.errno == errno.ENOSPC: + # Reduce disk max space to current - 256 MiB, cap cache and retry + with self._cache_lock: + sm = sum([st.size() for st in self._cache]) + self.cache_max = max((256 * 1024 * 1024), sm - (256 * 1024 * 1024)) + elif e.errno == errno.ENODEV: + _logger.error("Unable to use disk cache: The underlying filesystem does not support memory mapping.") + except Exception as e: + pass + finally: + # Check if we should evict things from the cache. Either + # because we added a new thing or there was an error and + # we possibly adjusted the limits down, so we might need + # to push something out. + self.cap_cache() + + try: + # Only gets here if there was an error the first time. The + # exception handler adjusts limits downward in some cases + # to free up resources, which would make the operation + # succeed. + slot.set(blob) + except Exception as e: + # It failed again. Give up. + slot.set(None) + raise arvados.errors.KeepCacheError("Unable to save block %s to disk cache: %s" % (slot.locator, e)) + + self.cap_cache() + class Counter(object): def __init__(self, v=0): self._lk = threading.Lock() @@ -1170,8 +1315,7 @@ class KeepClient(object): return blob finally: if slot is not None: - slot.set(blob) - self.block_cache.cap_cache() + self.block_cache.set(slot, blob) # Q: Including 403 is necessary for the Keep tests to continue # passing, but maybe they should expect KeepReadError instead?