KSM_overcommit: dynamic reserve calculation

Hi,

we are back with new features of KSM_overcommit test:
* NEW: guest_reserve and host_reserve are now calculated based on used memory
* NEW: tmpfs reserve is also evaluated to fit the overhead
* NEW: VM alive check during split_guest()
* FIX: In function split_guest() we used incorrect session
* MOD: Increase number of VNC ports

host_reserve:
* Still possible to set this value in cfg
* Calculation:
1) host_reserve = Available memory + minimal guest(128MB)
2) host_reserve += number of guests * 64

guest_reserve:
* Still possible to set this value in cfg
* Calculation:
1) guest_reserve = 256
2) guest_reserve += guest memory * constant
... where constant represents (TMPFS overhead and multiplicative OS memory consumption) per MB

Tested by ldoktor and jzupka on various HW (2GB-36GB host memory).

Best regards,
Lukáš Doktor

index 6bc7987..1d83120 100755
@@ -396,7 +396,7 @@ class VM:
 
             # Find available VNC port, if needed
             if params.get("display") == "vnc":
-                self.vnc_port = kvm_utils.find_free_port(5900, 6000)
+                self.vnc_port = kvm_utils.find_free_port(5900, 6100)
 
             # Find random UUID if specified 'uuid = random' in config file
             if params.get("uuid") == "random":

index 1036893..227745a 100755
@@ -8,10 +8,12 @@ Auxiliary script used to allocate memory on guests.
 """
 
 
-import os, array, sys, struct, random, copy, inspect, tempfile, datetime
+import os, array, sys, struct, random, copy, inspect, tempfile, datetime, math
 
 PAGE_SIZE = 4096 # machine page size
 
+TMPFS_OVERHEAD = 0.0022 # overhead on 1MB of write data 
+
 
 class MemFill(object):
     """

@@ -32,7 +34,8 @@ class MemFill(object):
 
         self.tmpdp = tempfile.mkdtemp()
         ret_code = os.system("mount -o size=%dM tmpfs %s -t tmpfs" %
-                             ((mem + 25), self.tmpdp))
+                             ((mem+math.ceil(mem*TMPFS_OVERHEAD)), 
+                             self.tmpdp))
         if ret_code != 0:
             if os.getuid() != 0:
                 print ("FAIL: Unable to mount tmpfs "

@@ -42,7 +45,7 @@ class MemFill(object):
         else:
             self.f = tempfile.TemporaryFile(prefix='mem', dir=self.tmpdp)
             self.allocate_by = 'L'
-            self.npages = (mem * 1024 * 1024) / PAGE_SIZE
+            self.npages = ((mem * 1024 * 1024) / PAGE_SIZE)
             self.random_key = random_key
             self.static_value = static_value
             print "PASS: Initialization"

@@ -83,7 +86,7 @@ class MemFill(object):
         @return: return array of bytes size PAGE_SIZE.
         """
         a = array.array("B")
-        for i in range(PAGE_SIZE / a.itemsize):
+        for i in range((PAGE_SIZE / a.itemsize)):
             try:
                 a.append(value)
             except:

index ee83ac2..d3a5982 100644
@@ -302,9 +302,9 @@ variants:
         ksm_overcommit_ratio = 3
         # Max paralel runs machine
         ksm_parallel_ratio = 4
-        # Host memory reserve
-        ksm_host_reserve = 512
-        ksm_guest_reserve = 1024
+        # Host memory reserve (default - best fit for used mem)
+        # ksm_host_reserve = 512
+        # ksm_guest_reserve = 1024
         variants:
             - ksm_serial:
                 ksm_mode = "serial"

Ok guys, thanks for working on this automatic evaluation of reserved memory for host and guest!

Ok, this sounds good, I'll give my round of testing on it after you guys re-send the patch to me with the small comments below addressed.

Also, I noticed you guys are not using the full awesomeness of git. Please read the mini tutorial I made so you can use git send-email.

http://autotest.kernel.org/wiki/GitWorkflow

I made comments on your patch using ------ so it's easier to identify that is a comment and what is the actual patch. Next time, please send the patch using git send-email.

index 6bc7987..1d83120 100755
@@ -396,7 +396,7 @@ class VM:
 
             # Find available VNC port, if needed
             if params.get("display") == "vnc":
-                self.vnc_port = kvm_utils.find_free_port(5900, 6000)
+                self.vnc_port = kvm_utils.find_free_port(5900, 6100)
 
             # Find random UUID if specified 'uuid = random' in config file
             if params.get("uuid") == "random":

index 1036893..227745a 100755
@@ -8,10 +8,12 @@ Auxiliary script used to allocate memory on guests.
 """
 
 
-import os, array, sys, struct, random, copy, inspect, tempfile, datetime
+import os, array, sys, struct, random, copy, inspect, tempfile, datetime, math
 
 PAGE_SIZE = 4096 # machine page size
 
+TMPFS_OVERHEAD = 0.0022 # overhead on 1MB of write data 
+
 
-------------------------------------------------------------------------------

Cool, how did you guys get to this constant?
-------------------------------------------------------------------------------

class MemFill(object):

     """

@@ -32,7 +34,8 @@ class MemFill(object):

self.tmpdp = tempfile.mkdtemp()
ret_code = os.system("mount -o size=%dM tmpfs %s -t tmpfs" %

-                             ((mem + 25), self.tmpdp))
+                             ((mem+math.ceil(mem*TMPFS_OVERHEAD)), 
+                             self.tmpdp))
         if ret_code != 0:
             if os.getuid() != 0:
                 print ("FAIL: Unable to mount tmpfs "

@@ -42,7 +45,7 @@ class MemFill(object):
else:
self.f = tempfile.TemporaryFile(prefix='mem', dir=self.tmpdp) self.allocate_by = 'L'
- self.npages = (mem * 1024 * 1024) / PAGE_SIZE
+ self.npages = ((mem * 1024 * 1024) / PAGE_SIZE)
self.random_key = random_key
self.static_value = static_value
print "PASS: Initialization"
@@ -83,7 +86,7 @@ class MemFill(object):

         @return: return array of bytes size PAGE_SIZE.
         """
         a = array.array("B")

- for i in range(PAGE_SIZE / a.itemsize):
+ for i in range((PAGE_SIZE / a.itemsize)):
try:
a.append(value)
except:

index 2dd46c4..31d5c61 100644
@@ -142,6 +142,10 @@ def run_ksm_overcommit(test, params, env):
         session = None
         vm = None
         for i in range(1, vmsc):
+            # Check VMs
+            for j in range(0, vmsc):
+                if not lvms[i].is_alive:
+                    raise error.TestFail("one of other VMs is death")
 
-------------------------------------------------------------------------------

The message above should be along the lines

e_msg = "VM %s died while executing static_random_fill on allocator loop" % i
-------------------------------------------------------------------------------
vm = lvms[i]
session = lsessions[i]
a_cmd = "mem.static_random_fill()"
@@ -154,6 +158,8 @@ def run_ksm_overcommit(test, params, env):
logging.debug("Watching host memory while filling vm %s memory",
vm.name)
while not out.startswith("PASS") and not out.startswith("FAIL"):
+ if not vm.is_alive():
+ raise error.TestFail("VM is death")

-------------------------------------------------------------------------------
The message above also should be along the lines

e_msg = "VM %s died while executing static_random_fill on allocator loop" % i
-------------------------------------------------------------------------------

free_mem = int(utils.read_from_meminfo("MemFree")) if (ksm_swap):
free_mem = (free_mem +
@@ -202,7 +208,7 @@ def run_ksm_overcommit(test, params, env):

# Verify last machine with randomly generated memory a_cmd = "mem.static_random_verify()"
- _execute_allocator(a_cmd, lvms[last_vm], session,
+ _execute_allocator(a_cmd, lvms[last_vm], lsessions[last_vm],

                            (mem / 200 * 50 * perf_ratio))
         logging.debug(kvm_test_utils.get_memory_info([lvms[last_vm]]))

@@ -338,12 +344,29 @@ def run_ksm_overcommit(test, params, env):

# Main test code
logging.info("Starting phase 0: Initialization")
+
# host_reserve: mem reserve kept for the host system to run

-    host_reserve = int(params.get("ksm_host_reserve", 512))
+    host_reserve = int(params.get("ksm_host_reserve", -1))
+    if (host_reserve == -1):
+        # default host_reserve = MemAvailable + one_minimal_guest(128MB)
+        # later we add 64MB per additional guest
+        host_reserve = ((utils.memtotal() - utils.read_from_meminfo("MemFree"))

+ / 1024 + 128)
+ # using default reserve
+ _host_reserve = 1

-------------------------------------------------------------------------------
Be consistent above and assign True instead of 1
-------------------------------------------------------------------------------

+ else:
+ _host_reserve = False
+
# guest_reserve: mem reserve kept to avoid guest OS to kill processes

-    guest_reserve = int(params.get("ksm_guest_reserve", 1024))
-    logging.debug("Memory reserved for host to run: %d", host_reserve)
-    logging.debug("Memory reserved for guest to run: %d", guest_reserve)
+    guest_reserve = int(params.get("ksm_guest_reserve", -1))
+    if (guest_reserve == -1):
+        # default guest_reserve = minimal_system_mem(256MB)
+        # later we add tmpfs overhead
+        guest_reserve = 256
+        # using default reserve
+        _guest_reserve = True

+ else:
+ _guest_reserve = False

max_vms = int(params.get("max_vms", 2)) overcommit = float(params.get("ksm_overcommit_ratio", 2.0))
@@ -355,6 +378,10 @@ def run_ksm_overcommit(test, params, env):

if (params['ksm_mode'] == "serial"):
max_alloc = vmsc
+ if _host_reserve:

+            # First round of additional guest reserves
+            host_reserve += vmsc * 64
+            _host_reserve = vmsc

     host_mem = (int(utils.memtotal()) / 1024 - host_reserve)

@@ -402,6 +429,10 @@ def run_ksm_overcommit(test, params, env):

             if mem - guest_reserve - 1 > 3100:
                 vmsc = int(math.ceil((host_mem * overcommit) /
                                      (3100 + guest_reserve)))
+                if _host_reserve:
+                    host_reserve += (vmsc - _host_reserve) * 64
+                    host_mem -= (vmsc - _host_reserve) * 64
+                    _host_reserve = vmsc

mem = int(math.floor(host_mem * overcommit / vmsc))

if os.popen("uname -i").readline().startswith("i386"):
@@ -410,8 +441,19 @@ def run_ksm_overcommit(test, params, env):

             if mem > 3100 - 64:
                 vmsc = int(math.ceil((host_mem * overcommit) /
                                      (3100 - 64.0)))
+                if _host_reserve:
+                    host_reserve += (vmsc - _host_reserve) * 64
+                    host_mem -= (vmsc - _host_reserve) * 64
+                    _host_reserve = vmsc

mem = int(math.floor(host_mem * overcommit / vmsc))

+ # 0.055 represents OS + TMPFS additional reserve per guest ram MB
+ if _guest_reserve:
+ guest_reserve += math.ceil(mem * 0.055)
+
+ logging.debug("Memory reserved for host to run: %d", host_reserve)
+ logging.debug("Memory reserved for guest to run: %d", guest_reserve)
+
logging.debug("Checking KSM status...") ksm_flag = 0
for line in os.popen('ksmctl info').readlines():

index ee83ac2..d3a5982 100644
@@ -302,9 +302,9 @@ variants:
         ksm_overcommit_ratio = 3
         # Max paralel runs machine
         ksm_parallel_ratio = 4
-        # Host memory reserve
-        ksm_host_reserve = 512
-        ksm_guest_reserve = 1024
+        # Host memory reserve (default - best fit for used mem)
+        # ksm_host_reserve = 512
+        # ksm_guest_reserve = 1024
         variants:
             - ksm_serial:
                 ksm_mode = "serial"