February 1, 2013 at 9:35 am
This is an interesting one.
Windows 2008r2SP1 EE
SQL 2008r2SP2CU1 EE
We have installed Cluster Service Update KB 2710487, which is supposed to fix inexplicable Cluster Service crashes with error 1359.
"The cluster service encountered an unexpected problem and will be shut down. The error code was '1359'"
The problem is that the Update increases CPU consumption a great deal. We've identified some interesting behavior:
•Install the patch on one Node, reboot the Node. CPU is normal (0) until you move a cluster group to that node
•With each Cluster Group Move a 1.25% increase in CPU is added to the Node, and it persists. Forever. Note that these are dedicated Nodes, and SQL Groups are complete inactive (no users, no databases)
•This CPU increase persists even after all Cluster Groups are moved off the Node
•3 processes clearly account for the total CPU increase. When comparing these processes to non-updated hosts we found each of their CPU usage has clearly increased, and each are in constant use. On non-patched hosts all 3 processes have little CPU usage, and are in use infrequently. Here's some metrics:
PROCESSES ON A NON-UPDATED NODE
ProcessCPUFrequency
------------------ --------- ------------------------------
Clussvc1Frequent but not steady
lssas0infrequent
Wmiprvse0infrequent
---
1Total Avg CPU
PROCESSES ON AN UPDATED NODE
ProcessCPUFrequency
------------------ --------- ------------------------------
Clussvc10Constant
lssas2Constant
Wmiprvse1Constant
---
13Total Avg CPU
•When a workload is placed on the Cluster Node, CPU consumption of these 3 processes increases. The worst we're seeing is a 15% CPU increase on each node, where each Node is hosting 4 active SQL Groups. The inactive Node in this Cluster has a steady 5% CPU load. It is hosting 4 SQL Groups with activity.
•CPU usage patterns by these 3 processes on Updated Nodes are the same regardless of Node CPU and Memory resources
MS tried to argue that this was not a significant increase... riiiiiiiiight. So after days and days of data collection and arguing with support, they've finally agreed this is a bug, and it will be fixed.
Unfortunately this is as far as we've gotten. We have no root cause, or time frame for the fix. We badly need to fix this 1359 error, but we can't afford a 15% increase in CPU.
One thing we've found is that Clusters with Updated Nodes have overly verbose Heartbeat chatter:
Cluster Update StatusPacket/Sec
--------------------------- ----------------
0 of 5 nodes Updated 2572
1 of 2 nodes Updated 5795
2 of 2 nodes Updated12008
4 of 4 nodes Updated 18865
Have any of you seen this issue or have thoughts on HB chattiness being a root cause for CPU load increase?
Viewing 0 posts
You must be logged in to reply to this topic. Login to reply