Performance management in virtualized environments is the topic of Gartner analyst Cameron Haight's post, Server Virtualization Performance. Haight notes that during a Gartner client gathering, he conducted an informal poll to see what part of the virtual server infrastructure caused clients the most performance problems. While individual VMs came out on top at 42%, Haight found that the underlying issue was very likely an application problem.
What Haight's clients possibly were experiencing is what we at ExtraHop have coined virtual packet loss, or VPL. This issue is very common in heavily virtualized environments and is usually brought about by an oversubscribed physical host running too many virtual machines. Although the VMs running on an oversubscribed host continue to operate, virtualized applications perform very poorly.
As a virtual machine gets swapped out, acknowledgments for packets it receives are delayed substantially. If the delay is too long, the sender may consider the packets lost, assume that the network is congested, and back off its transmission rate. To relieve this non-existent congestion, the sender will attempt multiple times to retransmit the packets, backing off every time. Because throughput is stalled, the network appears to be losing packets, even though no loss actually is occurring. This situation has a domino effect that results in slow throughput, longer data transfer rates, and sluggish application performance. And, as many of our customers have found, this perniciously intermittent phenomenon goes undetected by conventional monitoring and management tools.
ExtraHop's Application Delivery Assurance system helps numerous customers cope with VPL by pinpointing the precise root cause. Pulling packet data off the wire and analyzing it in real time, our system provides detailed metrics about when and where VPL is occurring. This advantage gives both the network and application teams essential information for capacity planning within virtual environments.
If you're interested in reading more about VPL, take a look at an article I wrote for Virtualization Journal: Virtual Packet Loss: The Silent Killer of Network Performance. I'd also be interested in hearing your thoughts on VPL in the comments below.
This is a companion discussion topic for the original entry at http://www.extrahop.com/post/blog/vm-virtual-packet-loss/