Identifying resource usage and performing migration to address performance issues in a Xen server virtualization environment
When Netflix faced an outage in 2011, they had to offer a 3% credit to their customers. The outage was due to a resource crunch in the Amazon cloud they were hosted in. It’s tough for a business to survive if such outages happen frequently.
A cloud hosting provider achieves economies of scale by sharing resources between multiple tenants in the cloud. These resources include computing, storage and network devices. But at times, the cloud provider may oversell the resources to achieve high-density or a single tenant can abuse the server resources. This can affect the performance of the cloud system.
A common performance metric in server virtualization systems is “CPU steal time”. It refers to the amount of time a VM is ready, but could not run due to other VMs competing for the CPU. Recently we were contacted by a cloud hosting provider whose customers were complaining of intermittent slowness and high CPU steal times.
Debugging the slowness issue
Steal time is the duration a VM waits for a real CPU cycle, while the hypervisor is servicing another VM. In this cloud solution, Xen hypervisor was used to manage the VMs. CPU stealing in Xen can happen due to:
- Resource-intensive applications taking up too much CPU in individual VMs.
- The cloud servers running short of resources.
To identify the reason for CPU stealing, we monitored the individual VMs in the cloud. Resource-intensive applications and frequency of high steal times were audited using ‘top’ utility. We noticed that CPU steal times of the slow VMs were continuously over 60%.
top - 10:26:21 up 45 days, 2:00, 2 users, load average: 8.01, 4.04, 2.46 Tasks: 496 total, 21 running, 475 sleeping, 0 stopped, 0 zombie Cpu(s): 29.8%us, 3.1%sy, 0.7%ni, 0.0%id, 0.0%wa, 0.0%hi, 0.0%si, 61.1%st
But adding more CPU for those VMs was not possible, as all the cloud servers were already near their physical limits.
Migrating the slow VMs
To improve speed and responsiveness, we decided to migrate the slow VMs to another host. We executed the following plan to accomplish this.
1.We configured a new host server with CentOS 6 and installed Xen hypervisor packages in it.
[root@Host02 ~]# yum install centos-release-xen
[root@Host02 ~]# yum install xen
2. We then updated the boot loader to load this new image and rebooted the machine.
[root@Host02 ~]# sh /usr/bin/grub-bootxen.sh Updating grub config
3. Once the Xen hypervisor started running, we migrated the slow VMs from the old host to the new server. The VMs were allotted enough CPU and resources to ensure that they ran without any hiccups.
[root@Host02 ~]# xm list Name ID Mem VCPUs State Time(s) Domain-0 0 8192 32 r----- 4589703.7 lhry2j7b1108p3 02 3048 3 -b---- 785333.0 res69fvhdx8h12 03 512 2 -b---- 84255.3 tqr01qycerunum 04 512 2 r----- 3381600.7
4. After the migration was done, we audited the resource usage in the VM nodes and found that the steal time was normal.
top - 11:15:01 up 2 days, 2:00, 2 users, load average: 0.01, 0.04, 0.06 Tasks: 96 total, 5 running, 91 sleeping, 0 stopped, 0 zombie Cpu(s): 0.0%us, 0.1%sy, 0.0%ni, 99.9%id, 0.0%wa, 0.0%hi, 0.0%si, 0.0%st
Monitoring script for Xen VMs
To avoid downtime and slowness issues in future, we configured a Perl script for monitoring the VMs in our Xen nodes. We installed the ‘Xentop’ utility to identify the processes that consume resources in each VM. The script would monitor the CPU usage in the VMs and send alert notification to our support team whenever there is a spike. Upon receiving the alert, we would debug and resolve it before it affected the customers.
Performance is a key requirement for any cloud hosting solution. Here we’ve covered how we scaled up a Xen cloud environment and did pro-active monitoring to ensure 100% uptime for all VMs. Bobcares helps web hosts, VPS providers and cloud providers deliver industry standard VPS services through custom configuration and preventive maintenance of virtualized systems.