Merge remote-tracking branch 'qubesos/pr/179'
* qubesos/pr/179: qmemman: request VMs balloon down with 16MB safety margin qmemman: clear "not responding" flags when VM require more memory qmemman: slightly improve logging qmemman: reformat code, especially comments
This commit is contained in:
commit
21760d8ff0
@ -40,12 +40,13 @@ class DomainState:
|
|||||||
self.memory_current = 0 # the current memory size
|
self.memory_current = 0 # the current memory size
|
||||||
self.memory_actual = None # the current memory allocation (what VM
|
self.memory_actual = None # the current memory allocation (what VM
|
||||||
# is using or can use at any time)
|
# is using or can use at any time)
|
||||||
self.memory_maximum = None #the maximum memory size
|
self.memory_maximum = None # the maximum memory size
|
||||||
self.mem_used = None #used memory, computed based on meminfo
|
self.mem_used = None # used memory, computed based on meminfo
|
||||||
self.id = id #domain id
|
self.id = id # domain id
|
||||||
self.last_target = 0 #the last memset target
|
self.last_target = 0 # the last memset target
|
||||||
self.no_progress = False #no react to memset
|
self.no_progress = False # no react to memset
|
||||||
self.slow_memset_react = False #slow react to memset (after few tries still above target)
|
self.slow_memset_react = False # slow react to memset (after few
|
||||||
|
# tries still above target)
|
||||||
|
|
||||||
def __repr__(self):
|
def __repr__(self):
|
||||||
return self.__dict__.__repr__()
|
return self.__dict__.__repr__()
|
||||||
@ -61,9 +62,11 @@ class SystemState(object):
|
|||||||
self.BALOON_DELAY = 0.1
|
self.BALOON_DELAY = 0.1
|
||||||
self.XEN_FREE_MEM_LEFT = 50*1024*1024
|
self.XEN_FREE_MEM_LEFT = 50*1024*1024
|
||||||
self.XEN_FREE_MEM_MIN = 25*1024*1024
|
self.XEN_FREE_MEM_MIN = 25*1024*1024
|
||||||
# Overhead of per-page Xen structures, taken from OpenStack nova/virt/xenapi/driver.py
|
# Overhead of per-page Xen structures, taken from OpenStack
|
||||||
|
# nova/virt/xenapi/driver.py
|
||||||
# see https://wiki.openstack.org/wiki/XenServer/Overhead
|
# see https://wiki.openstack.org/wiki/XenServer/Overhead
|
||||||
# we divide total and free physical memory by this to get "assignable" memory
|
# we divide total and free physical memory by this to get
|
||||||
|
# "assignable" memory
|
||||||
self.MEM_OVERHEAD_FACTOR = 1.0 / 1.00781
|
self.MEM_OVERHEAD_FACTOR = 1.0 / 1.00781
|
||||||
try:
|
try:
|
||||||
self.ALL_PHYS_MEM = int(self.xc.physinfo()['total_memory']*1024 * self.MEM_OVERHEAD_FACTOR)
|
self.ALL_PHYS_MEM = int(self.xc.physinfo()['total_memory']*1024 * self.MEM_OVERHEAD_FACTOR)
|
||||||
@ -108,7 +111,7 @@ class SystemState(object):
|
|||||||
xen_free, assigned_but_unused, self.domdict))
|
xen_free, assigned_but_unused, self.domdict))
|
||||||
return xen_free - assigned_but_unused
|
return xen_free - assigned_but_unused
|
||||||
|
|
||||||
#refresh information on memory assigned to all domains
|
# refresh information on memory assigned to all domains
|
||||||
def refresh_memactual(self):
|
def refresh_memactual(self):
|
||||||
for domain in self.xc.domain_getinfo():
|
for domain in self.xc.domain_getinfo():
|
||||||
id = str(domain['domid'])
|
id = str(domain['domid'])
|
||||||
@ -125,47 +128,49 @@ class SystemState(object):
|
|||||||
self.domdict[id].memory_maximum = int(self.domdict[id].memory_maximum)*1024
|
self.domdict[id].memory_maximum = int(self.domdict[id].memory_maximum)*1024
|
||||||
else:
|
else:
|
||||||
self.domdict[id].memory_maximum = self.ALL_PHYS_MEM
|
self.domdict[id].memory_maximum = self.ALL_PHYS_MEM
|
||||||
# the previous line used to be
|
# the previous line used to be
|
||||||
# self.domdict[id].memory_maximum = domain['maxmem_kb']*1024
|
# self.domdict[id].memory_maximum = domain[
|
||||||
# but domain['maxmem_kb'] changes in self.mem_set as well, and this results in
|
# 'maxmem_kb']*1024
|
||||||
# the memory never increasing
|
# but domain['maxmem_kb'] changes in self.mem_set as well,
|
||||||
# in fact, the only possible case of nonexisting memory/static-max is dom0
|
# and this results in the memory never increasing
|
||||||
# see #307
|
# in fact, the only possible case of nonexisting
|
||||||
|
# memory/static-max is dom0
|
||||||
|
# see #307
|
||||||
|
|
||||||
def clear_outdated_error_markers(self):
|
def clear_outdated_error_markers(self):
|
||||||
# Clear outdated errors
|
# Clear outdated errors
|
||||||
for i in self.domdict.keys():
|
for i in self.domdict.keys():
|
||||||
if self.domdict[i].slow_memset_react and \
|
# clear markers excluding VM from memory balance, if:
|
||||||
self.domdict[i].memory_actual <= self.domdict[i].last_target + self.XEN_FREE_MEM_LEFT/4:
|
# - VM have responded to previous request (with some safety margin)
|
||||||
dom_name = self.xs.read('', '/local/domain/%s/name' % str(i))
|
# - VM request more memory than it has assigned
|
||||||
if dom_name is not None:
|
# The second condition avoids starving a VM, even when there is
|
||||||
# TODO: report it somewhere, qubesd or elsewhere
|
# some free memory available
|
||||||
pass
|
if self.domdict[i].memory_actual <= \
|
||||||
|
self.domdict[i].last_target + self.XEN_FREE_MEM_LEFT/2 or \
|
||||||
|
self.domdict[i].memory_actual < \
|
||||||
|
qubes.qmemman.algo.prefmem(self.domdict[i]):
|
||||||
self.domdict[i].slow_memset_react = False
|
self.domdict[i].slow_memset_react = False
|
||||||
|
|
||||||
if self.domdict[i].no_progress and \
|
|
||||||
self.domdict[i].memory_actual <= self.domdict[i].last_target + self.XEN_FREE_MEM_LEFT/4:
|
|
||||||
dom_name = self.xs.read('', '/local/domain/%s/name' % str(i))
|
|
||||||
if dom_name is not None:
|
|
||||||
# TODO: report it somewhere, qubesd or elsewhere
|
|
||||||
pass
|
|
||||||
self.domdict[i].no_progress = False
|
self.domdict[i].no_progress = False
|
||||||
|
|
||||||
#the below works (and is fast), but then 'xm list' shows unchanged memory value
|
# the below works (and is fast), but then 'xm list' shows unchanged
|
||||||
|
# memory value
|
||||||
def mem_set(self, id, val):
|
def mem_set(self, id, val):
|
||||||
self.log.info('mem-set domain {} to {}'.format(id, val))
|
self.log.info('mem-set domain {} to {}'.format(id, val))
|
||||||
self.domdict[id].last_target = val
|
self.domdict[id].last_target = val
|
||||||
#can happen in the middle of domain shutdown
|
# can happen in the middle of domain shutdown
|
||||||
#apparently xc.lowlevel throws exceptions too
|
# apparently xc.lowlevel throws exceptions too
|
||||||
try:
|
try:
|
||||||
self.xc.domain_setmaxmem(int(id), int(val/1024) + 1024) # LIBXL_MAXMEM_CONSTANT=1024
|
self.xc.domain_setmaxmem(int(id), int(val/1024) + 1024) # LIBXL_MAXMEM_CONSTANT=1024
|
||||||
self.xc.domain_set_target_mem(int(id), int(val/1024))
|
self.xc.domain_set_target_mem(int(id), int(val / 1024))
|
||||||
except:
|
except:
|
||||||
pass
|
pass
|
||||||
self.xs.write('', '/local/domain/' + id + '/memory/target', str(int(val/1024)))
|
# VM sees about 16MB memory less, so adjust for it here - qmemman
|
||||||
|
# handle Xen view of memory
|
||||||
|
self.xs.write('', '/local/domain/' + id + '/memory/target',
|
||||||
|
str(int(val/1024 - 16 * 1024)))
|
||||||
|
|
||||||
# this is called at the end of ballooning, when we have Xen free mem already
|
# this is called at the end of ballooning, when we have Xen free mem already
|
||||||
# make sure that past mem_set will not decrease Xen free mem
|
# make sure that past mem_set will not decrease Xen free mem
|
||||||
def inhibit_balloon_up(self):
|
def inhibit_balloon_up(self):
|
||||||
self.log.debug('inhibit_balloon_up()')
|
self.log.debug('inhibit_balloon_up()')
|
||||||
for i in self.domdict.keys():
|
for i in self.domdict.keys():
|
||||||
@ -175,7 +180,8 @@ class SystemState(object):
|
|||||||
'Preventing balloon up to {}'.format(dom.last_target))
|
'Preventing balloon up to {}'.format(dom.last_target))
|
||||||
self.mem_set(i, dom.memory_actual)
|
self.mem_set(i, dom.memory_actual)
|
||||||
|
|
||||||
#perform memory ballooning, across all domains, to add "memsize" to Xen free memory
|
# perform memory ballooning, across all domains, to add "memsize" to Xen
|
||||||
|
# free memory
|
||||||
def do_balloon(self, memsize):
|
def do_balloon(self, memsize):
|
||||||
self.log.info('do_balloon(memsize={!r})'.format(memsize))
|
self.log.info('do_balloon(memsize={!r})'.format(memsize))
|
||||||
CHECK_PERIOD_S = 3
|
CHECK_PERIOD_S = 3
|
||||||
@ -213,7 +219,8 @@ class SystemState(object):
|
|||||||
if prev_memory_actual is not None:
|
if prev_memory_actual is not None:
|
||||||
for i in prev_memory_actual.keys():
|
for i in prev_memory_actual.keys():
|
||||||
if prev_memory_actual[i] == self.domdict[i].memory_actual:
|
if prev_memory_actual[i] == self.domdict[i].memory_actual:
|
||||||
#domain not responding to memset requests, remove it from donors
|
# domain not responding to memset requests, remove it
|
||||||
|
# from donors
|
||||||
self.domdict[i].no_progress = True
|
self.domdict[i].no_progress = True
|
||||||
self.log.info('domain {} stuck at {}'.format(i, self.domdict[i].memory_actual))
|
self.log.info('domain {} stuck at {}'.format(i, self.domdict[i].memory_actual))
|
||||||
memset_reqs = qubes.qmemman.algo.balloon(memsize + self.XEN_FREE_MEM_LEFT - xenfree, self.domdict)
|
memset_reqs = qubes.qmemman.algo.balloon(memsize + self.XEN_FREE_MEM_LEFT - xenfree, self.domdict)
|
||||||
@ -238,8 +245,8 @@ class SystemState(object):
|
|||||||
self.domdict[domid], untrusted_meminfo_key)
|
self.domdict[domid], untrusted_meminfo_key)
|
||||||
self.do_balance()
|
self.do_balance()
|
||||||
|
|
||||||
#is the computed balance request big enough ?
|
# is the computed balance request big enough ?
|
||||||
#so that we do not trash with small adjustments
|
# so that we do not trash with small adjustments
|
||||||
def is_balance_req_significant(self, memset_reqs, xenfree):
|
def is_balance_req_significant(self, memset_reqs, xenfree):
|
||||||
self.log.debug(
|
self.log.debug(
|
||||||
'is_balance_req_significant(memset_reqs={}, xenfree={})'.format(
|
'is_balance_req_significant(memset_reqs={}, xenfree={})'.format(
|
||||||
@ -261,7 +268,8 @@ class SystemState(object):
|
|||||||
total_memory_transfer += abs(memory_change)
|
total_memory_transfer += abs(memory_change)
|
||||||
pref = qubes.qmemman.algo.prefmem(self.domdict[dom])
|
pref = qubes.qmemman.algo.prefmem(self.domdict[dom])
|
||||||
|
|
||||||
if last_target > 0 and last_target < pref and memory_change > MIN_MEM_CHANGE_WHEN_UNDER_PREF:
|
if 0 < last_target < pref and \
|
||||||
|
memory_change > MIN_MEM_CHANGE_WHEN_UNDER_PREF:
|
||||||
self.log.info(
|
self.log.info(
|
||||||
'dom {} is below pref, allowing balance'.format(dom))
|
'dom {} is below pref, allowing balance'.format(dom))
|
||||||
return True
|
return True
|
||||||
@ -274,9 +282,14 @@ class SystemState(object):
|
|||||||
def print_stats(self, xenfree, memset_reqs):
|
def print_stats(self, xenfree, memset_reqs):
|
||||||
for i in self.domdict.keys():
|
for i in self.domdict.keys():
|
||||||
if self.domdict[i].mem_used is not None:
|
if self.domdict[i].mem_used is not None:
|
||||||
self.log.info('stat: dom {!r} act={} pref={}'.format(i,
|
self.log.info('stat: dom {!r} act={} pref={} last_target={}'
|
||||||
|
'{}{}'.format(i,
|
||||||
self.domdict[i].memory_actual,
|
self.domdict[i].memory_actual,
|
||||||
qubes.qmemman.algo.prefmem(self.domdict[i])))
|
qubes.qmemman.algo.prefmem(self.domdict[i]),
|
||||||
|
self.domdict[i].last_target,
|
||||||
|
' no_progress' if self.domdict[i].no_progress else '',
|
||||||
|
' slow_memset_react'
|
||||||
|
if self.domdict[i].slow_memset_react else ''))
|
||||||
|
|
||||||
self.log.info('stat: xenfree={} memset_reqs={}'.format(xenfree, memset_reqs))
|
self.log.info('stat: xenfree={} memset_reqs={}'.format(xenfree, memset_reqs))
|
||||||
|
|
||||||
@ -303,7 +316,7 @@ class SystemState(object):
|
|||||||
dom, mem = rq
|
dom, mem = rq
|
||||||
# Force to always have at least 0.9*self.XEN_FREE_MEM_LEFT (some
|
# Force to always have at least 0.9*self.XEN_FREE_MEM_LEFT (some
|
||||||
# margin for rounding errors). Before giving memory to
|
# margin for rounding errors). Before giving memory to
|
||||||
# domain, ensure that others have gived it back.
|
# domain, ensure that others have gave it back.
|
||||||
# If not - wait a little.
|
# If not - wait a little.
|
||||||
ntries = 5
|
ntries = 5
|
||||||
while self.get_free_xen_memory() - (mem - self.domdict[dom].memory_actual) < 0.9*self.XEN_FREE_MEM_LEFT:
|
while self.get_free_xen_memory() - (mem - self.domdict[dom].memory_actual) < 0.9*self.XEN_FREE_MEM_LEFT:
|
||||||
@ -318,11 +331,12 @@ class SystemState(object):
|
|||||||
for rq2 in memset_reqs:
|
for rq2 in memset_reqs:
|
||||||
dom2, mem2 = rq2
|
dom2, mem2 = rq2
|
||||||
if dom2 == dom:
|
if dom2 == dom:
|
||||||
# All donors have been procesed
|
# All donors have been processed
|
||||||
break
|
break
|
||||||
# allow some small margin
|
# allow some small margin
|
||||||
if self.domdict[dom2].memory_actual > self.domdict[dom2].last_target + self.XEN_FREE_MEM_LEFT/4:
|
if self.domdict[dom2].memory_actual > self.domdict[dom2].last_target + self.XEN_FREE_MEM_LEFT/4:
|
||||||
# VM didn't react to memory request at all, remove from donors
|
# VM didn't react to memory request at all,
|
||||||
|
# remove from donors
|
||||||
if prev_memactual[dom2] == self.domdict[dom2].memory_actual:
|
if prev_memactual[dom2] == self.domdict[dom2].memory_actual:
|
||||||
self.log.warning(
|
self.log.warning(
|
||||||
'dom {!r} didnt react to memory request'
|
'dom {!r} didnt react to memory request'
|
||||||
@ -331,10 +345,6 @@ class SystemState(object):
|
|||||||
self.domdict[dom2].memory_actual,
|
self.domdict[dom2].memory_actual,
|
||||||
mem2))
|
mem2))
|
||||||
self.domdict[dom2].no_progress = True
|
self.domdict[dom2].no_progress = True
|
||||||
dom_name = self.xs.read('', '/local/domain/%s/name' % str(dom2))
|
|
||||||
if dom_name is not None:
|
|
||||||
# TODO: report it somewhere, qubesd or elsewhere
|
|
||||||
pass
|
|
||||||
else:
|
else:
|
||||||
self.log.warning('dom {!r} still hold more'
|
self.log.warning('dom {!r} still hold more'
|
||||||
' memory than have assigned ({} > {})'
|
' memory than have assigned ({} > {})'
|
||||||
@ -342,10 +352,6 @@ class SystemState(object):
|
|||||||
self.domdict[dom2].memory_actual,
|
self.domdict[dom2].memory_actual,
|
||||||
mem2))
|
mem2))
|
||||||
self.domdict[dom2].slow_memset_react = True
|
self.domdict[dom2].slow_memset_react = True
|
||||||
dom_name = self.xs.read('', '/local/domain/%s/name' % str(dom2))
|
|
||||||
if dom_name is not None:
|
|
||||||
# TODO: report it somewhere, qubesd or elsewhere
|
|
||||||
pass
|
|
||||||
self.mem_set(dom, self.get_free_xen_memory() + self.domdict[dom].memory_actual - self.XEN_FREE_MEM_LEFT)
|
self.mem_set(dom, self.get_free_xen_memory() + self.domdict[dom].memory_actual - self.XEN_FREE_MEM_LEFT)
|
||||||
return
|
return
|
||||||
|
|
||||||
|
@ -26,14 +26,14 @@ import string
|
|||||||
# This are only defaults - can be overridden by QMemmanServer with values from
|
# This are only defaults - can be overridden by QMemmanServer with values from
|
||||||
# config file
|
# config file
|
||||||
CACHE_FACTOR = 1.3
|
CACHE_FACTOR = 1.3
|
||||||
MIN_PREFMEM = 200*1024*1024
|
MIN_PREFMEM = 200 * 1024 * 1024
|
||||||
DOM0_MEM_BOOST = 350*1024*1024
|
DOM0_MEM_BOOST = 350 * 1024 * 1024
|
||||||
|
|
||||||
|
|
||||||
log = logging.getLogger('qmemman.daemon.algo')
|
log = logging.getLogger('qmemman.daemon.algo')
|
||||||
|
|
||||||
#untrusted meminfo size is taken from xenstore key, thus its size is limited
|
|
||||||
#so splits do not require excessive memory
|
# untrusted meminfo size is taken from xenstore key, thus its size is limited
|
||||||
|
# so splits do not require excessive memory
|
||||||
def sanitize_and_parse_meminfo(untrusted_meminfo):
|
def sanitize_and_parse_meminfo(untrusted_meminfo):
|
||||||
if not untrusted_meminfo:
|
if not untrusted_meminfo:
|
||||||
return None
|
return None
|
||||||
@ -65,7 +65,6 @@ def sanitize_and_parse_meminfo(untrusted_meminfo):
|
|||||||
meminfo['MemFree'] - meminfo['Cached'] - meminfo['Buffers'] +
|
meminfo['MemFree'] - meminfo['Cached'] - meminfo['Buffers'] +
|
||||||
meminfo['SwapTotal'] - meminfo['SwapFree']) * 1024
|
meminfo['SwapTotal'] - meminfo['SwapFree']) * 1024
|
||||||
|
|
||||||
|
|
||||||
return None
|
return None
|
||||||
|
|
||||||
|
|
||||||
@ -89,7 +88,8 @@ def is_meminfo_suspicious(untrusted_meminfo):
|
|||||||
ret = True
|
ret = True
|
||||||
if untrusted_meminfo['MemTotal'] < \
|
if untrusted_meminfo['MemTotal'] < \
|
||||||
untrusted_meminfo['MemFree'] + \
|
untrusted_meminfo['MemFree'] + \
|
||||||
untrusted_meminfo['Cached'] + untrusted_meminfo['Buffers']:
|
untrusted_meminfo['Cached'] + untrusted_meminfo[
|
||||||
|
'Buffers']:
|
||||||
ret = True
|
ret = True
|
||||||
# we could also impose some limits on all the above values
|
# we could also impose some limits on all the above values
|
||||||
# but it has little purpose - all the domain can gain by passing e.g.
|
# but it has little purpose - all the domain can gain by passing e.g.
|
||||||
@ -97,7 +97,8 @@ def is_meminfo_suspicious(untrusted_meminfo):
|
|||||||
# it can be achieved with legal values, too, and it will not allow to
|
# it can be achieved with legal values, too, and it will not allow to
|
||||||
# starve existing domains, by design
|
# starve existing domains, by design
|
||||||
if ret:
|
if ret:
|
||||||
log.warning('suspicious meminfo untrusted_meminfo={!r}'.format(untrusted_meminfo))
|
log.warning('suspicious meminfo untrusted_meminfo={!r}'.format(
|
||||||
|
untrusted_meminfo))
|
||||||
return ret
|
return ret
|
||||||
|
|
||||||
|
|
||||||
@ -107,20 +108,26 @@ def refresh_meminfo_for_domain(domain, untrusted_xenstore_key):
|
|||||||
|
|
||||||
|
|
||||||
def prefmem(domain):
|
def prefmem(domain):
|
||||||
#dom0 is special, as it must have large cache, for vbds. Thus, give it a special boost
|
# dom0 is special, as it must have large cache, for vbds. Thus, give it
|
||||||
|
# a special boost
|
||||||
if domain.id == '0':
|
if domain.id == '0':
|
||||||
return min(domain.mem_used*CACHE_FACTOR + DOM0_MEM_BOOST, domain.memory_maximum)
|
return min(domain.mem_used * CACHE_FACTOR + DOM0_MEM_BOOST,
|
||||||
return max(min(domain.mem_used*CACHE_FACTOR, domain.memory_maximum), MIN_PREFMEM)
|
domain.memory_maximum)
|
||||||
|
return max(min(domain.mem_used * CACHE_FACTOR, domain.memory_maximum),
|
||||||
|
MIN_PREFMEM)
|
||||||
|
|
||||||
|
|
||||||
def memory_needed(domain):
|
def memory_needed(domain):
|
||||||
#do not change
|
# do not change
|
||||||
#in balance(), "distribute total_available_memory proportionally to mempref" relies on this exact formula
|
# in balance(), "distribute total_available_memory proportionally to
|
||||||
|
# mempref" relies on this exact formula
|
||||||
ret = prefmem(domain) - domain.memory_actual
|
ret = prefmem(domain) - domain.memory_actual
|
||||||
return ret
|
return ret
|
||||||
|
|
||||||
#prepare list of (domain, memory_target) pairs that need to be passed
|
|
||||||
#to "xm memset" equivalent in order to obtain "memsize" of memory
|
# prepare list of (domain, memory_target) pairs that need to be passed
|
||||||
#return empty list when the request cannot be satisfied
|
# to "xm memset" equivalent in order to obtain "memsize" of memory
|
||||||
|
# return empty list when the request cannot be satisfied
|
||||||
def balloon(memsize, domain_dictionary):
|
def balloon(memsize, domain_dictionary):
|
||||||
log.debug('balloon(memsize={!r}, domain_dictionary={!r})'.format(
|
log.debug('balloon(memsize={!r}, domain_dictionary={!r})'.format(
|
||||||
memsize, domain_dictionary))
|
memsize, domain_dictionary))
|
||||||
@ -137,26 +144,30 @@ def balloon(memsize, domain_dictionary):
|
|||||||
if need < 0:
|
if need < 0:
|
||||||
log.info('balloon: dom {} has actual memory {}'.format(i,
|
log.info('balloon: dom {} has actual memory {}'.format(i,
|
||||||
domain_dictionary[i].memory_actual))
|
domain_dictionary[i].memory_actual))
|
||||||
donors.append((i,-need))
|
donors.append((i, -need))
|
||||||
available-=need
|
available -= need
|
||||||
|
|
||||||
log.info('req={} avail={} donors={!r}'.format(memsize, available, donors))
|
log.info('req={} avail={} donors={!r}'.format(memsize, available, donors))
|
||||||
|
|
||||||
if available<memsize:
|
if available < memsize:
|
||||||
return ()
|
return ()
|
||||||
scale = 1.0*memsize/available
|
scale = 1.0 * memsize / available
|
||||||
for donors_iter in donors:
|
for donors_iter in donors:
|
||||||
id, mem = donors_iter
|
dom_id, mem = donors_iter
|
||||||
memborrowed = mem*scale*REQ_SAFETY_NET_FACTOR
|
memborrowed = mem * scale * REQ_SAFETY_NET_FACTOR
|
||||||
log.info('borrow {} from {}'.format(memborrowed, id))
|
log.info('borrow {} from {}'.format(memborrowed, dom_id))
|
||||||
memtarget = int(domain_dictionary[id].memory_actual - memborrowed)
|
memtarget = int(domain_dictionary[dom_id].memory_actual - memborrowed)
|
||||||
request.append((id, memtarget))
|
request.append((dom_id, memtarget))
|
||||||
return request
|
return request
|
||||||
# REQ_SAFETY_NET_FACTOR is a bit greater that 1. So that if the domain yields a bit less than requested, due
|
|
||||||
# to e.g. rounding errors, we will not get stuck. The surplus will return to the VM during "balance" call.
|
|
||||||
|
|
||||||
|
|
||||||
#redistribute positive "total_available_memory" of memory between domains, proportionally to prefmem
|
# REQ_SAFETY_NET_FACTOR is a bit greater that 1. So that if the domain
|
||||||
|
# yields a bit less than requested, due to e.g. rounding errors, we will not
|
||||||
|
# get stuck. The surplus will return to the VM during "balance" call.
|
||||||
|
|
||||||
|
|
||||||
|
# redistribute positive "total_available_memory" of memory between domains,
|
||||||
|
# proportionally to prefmem
|
||||||
def balance_when_enough_memory(domain_dictionary,
|
def balance_when_enough_memory(domain_dictionary,
|
||||||
xen_free_memory, total_mem_pref, total_available_memory):
|
xen_free_memory, total_mem_pref, total_available_memory):
|
||||||
log.info('balance_when_enough_memory(xen_free_memory={!r}, '
|
log.info('balance_when_enough_memory(xen_free_memory={!r}, '
|
||||||
@ -172,20 +183,21 @@ def balance_when_enough_memory(domain_dictionary,
|
|||||||
continue
|
continue
|
||||||
if domain_dictionary[i].no_progress:
|
if domain_dictionary[i].no_progress:
|
||||||
continue
|
continue
|
||||||
#distribute total_available_memory proportionally to mempref
|
# distribute total_available_memory proportionally to mempref
|
||||||
scale = 1.0*prefmem(domain_dictionary[i])/total_mem_pref
|
scale = 1.0 * prefmem(domain_dictionary[i]) / total_mem_pref
|
||||||
target_nonint = prefmem(domain_dictionary[i]) + scale*total_available_memory
|
target_nonint = prefmem(
|
||||||
#prevent rounding errors
|
domain_dictionary[i]) + scale * total_available_memory
|
||||||
target = int(0.999*target_nonint)
|
# prevent rounding errors
|
||||||
#do not try to give more memory than static max
|
target = int(0.999 * target_nonint)
|
||||||
|
# do not try to give more memory than static max
|
||||||
if target > domain_dictionary[i].memory_maximum:
|
if target > domain_dictionary[i].memory_maximum:
|
||||||
left_memory += target-domain_dictionary[i].memory_maximum
|
left_memory += target - domain_dictionary[i].memory_maximum
|
||||||
target = domain_dictionary[i].memory_maximum
|
target = domain_dictionary[i].memory_maximum
|
||||||
else:
|
else:
|
||||||
# count domains which can accept more memory
|
# count domains which can accept more memory
|
||||||
acceptors_count += 1
|
acceptors_count += 1
|
||||||
target_memory[i] = target
|
target_memory[i] = target
|
||||||
# distribute left memory across all acceptors
|
# distribute left memory across all acceptors
|
||||||
while left_memory > 0 and acceptors_count > 0:
|
while left_memory > 0 and acceptors_count > 0:
|
||||||
log.info('left_memory={} acceptors_count={}'.format(
|
log.info('left_memory={} acceptors_count={}'.format(
|
||||||
left_memory, acceptors_count))
|
left_memory, acceptors_count))
|
||||||
@ -195,9 +207,10 @@ def balance_when_enough_memory(domain_dictionary,
|
|||||||
for i in target_memory.keys():
|
for i in target_memory.keys():
|
||||||
target = target_memory[i]
|
target = target_memory[i]
|
||||||
if target < domain_dictionary[i].memory_maximum:
|
if target < domain_dictionary[i].memory_maximum:
|
||||||
memory_bonus = int(0.999*(left_memory/acceptors_count))
|
memory_bonus = int(0.999 * (left_memory / acceptors_count))
|
||||||
if target+memory_bonus >= domain_dictionary[i].memory_maximum:
|
if target + memory_bonus >= domain_dictionary[i].memory_maximum:
|
||||||
new_left_memory += target+memory_bonus - domain_dictionary[i].memory_maximum
|
new_left_memory += target + memory_bonus - \
|
||||||
|
domain_dictionary[i].memory_maximum
|
||||||
target = domain_dictionary[i].memory_maximum
|
target = domain_dictionary[i].memory_maximum
|
||||||
new_acceptors_count -= 1
|
new_acceptors_count -= 1
|
||||||
else:
|
else:
|
||||||
@ -205,26 +218,28 @@ def balance_when_enough_memory(domain_dictionary,
|
|||||||
target_memory[i] = target
|
target_memory[i] = target
|
||||||
left_memory = new_left_memory
|
left_memory = new_left_memory
|
||||||
acceptors_count = new_acceptors_count
|
acceptors_count = new_acceptors_count
|
||||||
# split target_memory dictionary to donors and acceptors
|
# split target_memory dictionary to donors and acceptors
|
||||||
# this is needed to first get memory from donors and only then give it to acceptors
|
# this is needed to first get memory from donors and only then give it
|
||||||
|
# to acceptors
|
||||||
donors_rq = list()
|
donors_rq = list()
|
||||||
acceptors_rq = list()
|
acceptors_rq = list()
|
||||||
for i in target_memory.keys():
|
for i in target_memory.keys():
|
||||||
target = target_memory[i]
|
target = target_memory[i]
|
||||||
if (target < domain_dictionary[i].memory_actual):
|
if target < domain_dictionary[i].memory_actual:
|
||||||
donors_rq.append((i, target))
|
donors_rq.append((i, target))
|
||||||
else:
|
else:
|
||||||
acceptors_rq.append((i, target))
|
acceptors_rq.append((i, target))
|
||||||
|
|
||||||
# print 'balance(enough): xen_free_memory=', xen_free_memory, 'requests:', donors_rq + acceptors_rq
|
# print 'balance(enough): xen_free_memory=', xen_free_memory, \
|
||||||
|
# 'requests:', donors_rq + acceptors_rq
|
||||||
return donors_rq + acceptors_rq
|
return donors_rq + acceptors_rq
|
||||||
|
|
||||||
|
|
||||||
#when not enough mem to make everyone be above prefmem, make donors be at prefmem, and
|
# when not enough mem to make everyone be above prefmem, make donors be at
|
||||||
#redistribute anything left between acceptors
|
# prefmem, and redistribute anything left between acceptors
|
||||||
def balance_when_low_on_memory(domain_dictionary,
|
def balance_when_low_on_memory(domain_dictionary,
|
||||||
xen_free_memory, total_mem_pref_acceptors, donors, acceptors):
|
xen_free_memory, total_mem_pref_acceptors, donors, acceptors):
|
||||||
log.debug('balance_when_low_on_memory(xen_free_memory={!r}, '
|
log.info('balance_when_low_on_memory(xen_free_memory={!r}, '
|
||||||
'total_mem_pref_acceptors={!r}, donors={!r}, acceptors={!r})'.format(
|
'total_mem_pref_acceptors={!r}, donors={!r}, acceptors={!r})'.format(
|
||||||
xen_free_memory, total_mem_pref_acceptors, donors, acceptors))
|
xen_free_memory, total_mem_pref_acceptors, donors, acceptors))
|
||||||
donors_rq = list()
|
donors_rq = list()
|
||||||
@ -232,55 +247,63 @@ def balance_when_low_on_memory(domain_dictionary,
|
|||||||
squeezed_mem = xen_free_memory
|
squeezed_mem = xen_free_memory
|
||||||
for i in donors:
|
for i in donors:
|
||||||
avail = -memory_needed(domain_dictionary[i])
|
avail = -memory_needed(domain_dictionary[i])
|
||||||
if avail < 10*1024*1024:
|
if avail < 10 * 1024 * 1024:
|
||||||
#probably we have already tried making it exactly at prefmem, give up
|
# probably we have already tried making it exactly at prefmem,
|
||||||
|
# give up
|
||||||
continue
|
continue
|
||||||
squeezed_mem -= avail
|
squeezed_mem -= avail
|
||||||
donors_rq.append((i, prefmem(domain_dictionary[i])))
|
donors_rq.append((i, prefmem(domain_dictionary[i])))
|
||||||
#the below can happen if initially xen free memory is below 50M
|
# the below can happen if initially xen free memory is below 50M
|
||||||
if squeezed_mem < 0:
|
if squeezed_mem < 0:
|
||||||
return donors_rq
|
return donors_rq
|
||||||
for i in acceptors:
|
for i in acceptors:
|
||||||
scale = 1.0*prefmem(domain_dictionary[i])/total_mem_pref_acceptors
|
scale = 1.0 * prefmem(domain_dictionary[i]) / total_mem_pref_acceptors
|
||||||
target_nonint = domain_dictionary[i].memory_actual + scale*squeezed_mem
|
target_nonint = \
|
||||||
#do not try to give more memory than static max
|
domain_dictionary[i].memory_actual + scale * squeezed_mem
|
||||||
target = min(int(0.999*target_nonint), domain_dictionary[i].memory_maximum)
|
# do not try to give more memory than static max
|
||||||
|
target = \
|
||||||
|
min(int(0.999 * target_nonint), domain_dictionary[i].memory_maximum)
|
||||||
acceptors_rq.append((i, target))
|
acceptors_rq.append((i, target))
|
||||||
# print 'balance(low): xen_free_memory=', xen_free_memory, 'requests:', donors_rq + acceptors_rq
|
# print 'balance(low): xen_free_memory=', xen_free_memory, 'requests:',
|
||||||
|
# donors_rq + acceptors_rq
|
||||||
return donors_rq + acceptors_rq
|
return donors_rq + acceptors_rq
|
||||||
|
|
||||||
|
|
||||||
#redistribute memory across domains
|
# redistribute memory across domains
|
||||||
#called when one of domains update its 'meminfo' xenstore key
|
# called when one of domains update its 'meminfo' xenstore key
|
||||||
#return the list of (domain, memory_target) pairs to be passed to
|
# return the list of (domain, memory_target) pairs to be passed to
|
||||||
#"xm memset" equivalent
|
# "xm memset" equivalent
|
||||||
def balance(xen_free_memory, domain_dictionary):
|
def balance(xen_free_memory, domain_dictionary):
|
||||||
log.debug('balance(xen_free_memory={!r}, domain_dictionary={!r})'.format(
|
log.debug('balance(xen_free_memory={!r}, domain_dictionary={!r})'.format(
|
||||||
xen_free_memory, domain_dictionary))
|
xen_free_memory, domain_dictionary))
|
||||||
|
|
||||||
#sum of all memory requirements - in other words, the difference between
|
# sum of all memory requirements - in other words, the difference between
|
||||||
#memory required to be added to domains (acceptors) to make them be at their
|
# memory required to be added to domains (acceptors) to make them be
|
||||||
#preferred memory, and memory that can be taken from domains (donors) that
|
# at their preferred memory, and memory that can be taken from domains
|
||||||
#can provide memory. So, it can be negative when plenty of memory.
|
# (donors) that can provide memory. So, it can be negative when plenty
|
||||||
|
# of memory.
|
||||||
total_memory_needed = 0
|
total_memory_needed = 0
|
||||||
|
|
||||||
#sum of memory preferences of all domains
|
# sum of memory preferences of all domains
|
||||||
total_mem_pref = 0
|
total_mem_pref = 0
|
||||||
|
|
||||||
#sum of memory preferences of all domains that require more memory
|
# sum of memory preferences of all domains that require more memory
|
||||||
total_mem_pref_acceptors = 0
|
total_mem_pref_acceptors = 0
|
||||||
|
|
||||||
donors = list() # domains that can yield memory
|
donors = list() # domains that can yield memory
|
||||||
acceptors = list() # domains that require more memory
|
acceptors = list() # domains that require more memory
|
||||||
#pass 1: compute the above "total" values
|
# pass 1: compute the above "total" values
|
||||||
for i in domain_dictionary.keys():
|
for i in domain_dictionary.keys():
|
||||||
if domain_dictionary[i].mem_used is None:
|
if domain_dictionary[i].mem_used is None:
|
||||||
continue
|
continue
|
||||||
if domain_dictionary[i].no_progress:
|
if domain_dictionary[i].no_progress:
|
||||||
continue
|
continue
|
||||||
need = memory_needed(domain_dictionary[i])
|
need = memory_needed(domain_dictionary[i])
|
||||||
# print 'domain' , i, 'act/pref', domain_dictionary[i].memory_actual, prefmem(domain_dictionary[i]), 'need=', need
|
# print 'domain' , i, 'act/pref', \
|
||||||
if need < 0 or domain_dictionary[i].memory_actual >= domain_dictionary[i].memory_maximum:
|
# domain_dictionary[i].memory_actual, prefmem(domain_dictionary[i]), \
|
||||||
|
# 'need=', need
|
||||||
|
if need < 0 or domain_dictionary[i].memory_actual >= \
|
||||||
|
domain_dictionary[i].memory_maximum:
|
||||||
donors.append(i)
|
donors.append(i)
|
||||||
else:
|
else:
|
||||||
acceptors.append(i)
|
acceptors.append(i)
|
||||||
@ -290,6 +313,8 @@ def balance(xen_free_memory, domain_dictionary):
|
|||||||
|
|
||||||
total_available_memory = xen_free_memory - total_memory_needed
|
total_available_memory = xen_free_memory - total_memory_needed
|
||||||
if total_available_memory > 0:
|
if total_available_memory > 0:
|
||||||
return balance_when_enough_memory(domain_dictionary, xen_free_memory, total_mem_pref, total_available_memory)
|
return balance_when_enough_memory(domain_dictionary, xen_free_memory,
|
||||||
|
total_mem_pref, total_available_memory)
|
||||||
else:
|
else:
|
||||||
return balance_when_low_on_memory(domain_dictionary, xen_free_memory, total_mem_pref_acceptors, donors, acceptors)
|
return balance_when_low_on_memory(domain_dictionary, xen_free_memory,
|
||||||
|
total_mem_pref_acceptors, donors, acceptors)
|
||||||
|
Loading…
Reference in New Issue
Block a user