Archive for July, 2009
Our VMWare ESX server does us a great job.
Running on an IBM X3650 HW, with 24GB RAM and 2×4 cores, it can simultaneously run up to 25 virtual machines, each VM is configured with around ~1.5 GB of RAM.
After reaching the 25 running VMs mark, we started noticing increasing sluggishness when additional VMs were turned on.
Of course, we did the trivial stuff of making sure that all screen savers are disabled, antivirus agents are not correlated to run at the same point in time, and making sure that all of the VMs are running the latest VMWare tools agent.
It was time to dig in deeper to find out where is the bottleneck we came across.
Someone told me that the stats that the reliability of the performance indicators that the graphic VI console shows is questionable and it’s recommended using the terminal utilities.So, I SHHed to the service console VM and ran the top utility. Immediately, I understood that what I’m actually doing is surveying the service console VM processes, rather than the overall ESX hypervisor activity. A quick dig up made me realize that the hypervisor is visible through the esxtop command, which is also executed from within the service console VM.
even for those of you that knows your way through the output of top and linux’s sysstat package, the data shown by esxtop is rather cryptic.
This great esxtop tutorial did me a great service with understanding the esxtop output.
I started more than 30 machines to reproduce the problem, and quickly went through the list of usual suspects: CPU, memory and IO:
I’ve verified that it’s not a CPU problem since the “CPU load average” was around 0.2. and PCPU was much the same.
Then I’ve switched to the memory display and verified that it’s not a physical memory issue. I saw the “high state” marker which was a good sign + there were almost 17GB ursvd (unreserved memory) in the VMKMEM/MB line.
SWAP (~3GB) seemed OK.
VMWare’s ballooning and memory sharing does miracles in broad day light.
I didn’t see any queues forming. read/write rates seemed pretty low.
So, the 25 VMs performance limit will remain a mystery until I’ll have proper time to analyze it more throughly, or even better, I’ll find someone from IT to do that for me.
Getting it right the first time
What happens when customers are experiencing problems with you application in production? The customer would send you the various logs artifacts and, ideally, you should be able to diagnose the problem and provide a resolution. If you find yourself sending the customer back and forth in an effort to gather additional types of log artifacts and system information, then you are, must likely, doing something wrong.
Who should be helping you
If you deploy your application on top of a application server platform, like Websphere Application Server (WAS) in my case, the platform should be assisting with automatic logs generation and collection. Our development team has been increasingly relying on such services provided by WAS, like: FFDC, WAS Collector, hung threads detection. All of which honorably earned their production stripes and badges.
One new serviceability artifact that I have long ago really wanted to have in production was the verbose GC, this feature records the JVM garbage collection activity over time, providing insight for resolving issues such as: stop-the-world performance freezes, memory leaks, native heap corruption, etc.
Until today, I was reluctant to enable the verbose GC in production, since I believed that there’s no way to direct the verbose GC output from the native stder (default) to a rotating dedicated file, not doing so might lead to files larger than 2GB (a problem on some file systems), or would cause the system to run out of disk space. I was assuming that the performance implications would be negligible, but still, you have to be extra prudent when it comes to live customers environments.
A trigger for action
Last week I had an issue with a WAS component, after opening a ticket with Websphere support, I was asked to reproduce the scenario in order to generate verbose GC output, I decided that enough is enough! I’m gonna look into the GC output file rollover issue again and see what solutions exist, what the community have to say about it, or whether there might be some other custom solution (with the Apache web server, for example, the file rolling is handled by an external process into which the log output is redirected, the process then does the rolling files management itself).
Following a quick search, I was happy to find that the IBM JVM offers a rolling over verbose GC. I quickly found additional hands on reports, Chris Bailey published verbose GC performance impact results that reassured my gut feeling about any performance impact being a non issue.
Here’s the syntax: (quoting the IBM Java 6 diagnostics guide):
Causes -verbose:gc output to be written to the specified file. If the file cannot be found, -verbose:gc tries to create the file, and then continues as normal if it is successful. If it cannot create the file (for example, if an invalid filename is passed into the command), it redirects the output to stderr.
If you specify and the -verbose:gc output is redirected to X files, each containing Y GC cycles.
- I don’t like having to specify the entire path for the file files, the default path should have been the server’s logs directory, or the CWD (CWD is the profile’s directory I believe).
- Rollover threshold parameter – I would rather be specifying it in units of max MBs instead of in units of the number of GC cycles entries. I’ve empirically found that 1MB of verbose GC log translates to ~700 GC cycle entries (YMMV).
- Good enough. I’ll start doing the preparations to put this into production.