1. © 2004 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice Understanding Virtualization Overhead to Optimize VM Mechanisms and Configurations Fabricio Benevenuto (Souza), Jose Renato Santos, Yoshio Turner, G. (John) Janakiraman ISSL, HP Labs

2. 2 Xen background IDD accesses physical device directly Guest access physical device indirectly through a virtual device connected to the IDD I/O Driver Domain Guest Domain I/O Channel peth0 Bridge vif1vif2 Virtual Machine Monitor Xen is an open source virtual machine monitor −Uses paravirtualization for better performance (requires OS modifications) Xen I/O model NIC

3. 3 Virtual environment configuration problem How many VMs can a given system support? How are the CPUs assigned among the VMs? How are the devices mapped among the VMs?  Understanding virtualization overheads is critical for optimal virtual environment configuration IDD CPU Guest NIC CPU IDD CPU Guest NIC CPU IDD Guest CPU Guest NIC CPU Guest IDD

4. 4 Focus of the internship Quantify virtualization overhead for optimal assignment of VMs to physical hosts −Characterize overhead of network I/O processing Sensitivity to application traffic profile Sensitivity to IDD scaling Sensitivity to device assignments Quantify performance interference among VMs caused by sharing physical resources not explicitly managed by the VMM −e.g., VM usage of processor caches and TLB

5. 5 Experimental setup 4-way SMP (Proliant DL-580) −Two gigE NICs (Intel E1000) connected to two client machines Benchmarks −TCP stream benchmark: single connection −Linux kernel compilation Used Xenoprof to measure CPU utilization, cache and TLB misses −Uses hardware counters to measure events such as cache misses, TLB misses, and clock cycles

6. 6 I/O processing overhead in Xen: Single NIC 3 to 5 times higher CPU utilization (IDD + guest) for same throughput compared to Linux What CPU allocation is needed to support I/O processing?

7. 7 What does IDD overhead depend on? IDD overhead is sensitive to packet rate – can be approximated by: U IDD = C 1 packets/s + C 2 acks/s CPU allocation can be determined using application’s traffic profile IDD CPU can be the limiter for I/O throughput

8. 8 Scaling IDD’s I/O capacity Can multiple CPUs in the IDD improve I/O throughput (e.g., with multiple NICs)? Assigning 1 CPU to each NIC degrades performance −Currently, I/O throughput is limited by the capacity of one IDD CPU Points to a problem in Xen – needs further investigation Alternative: an IDD for each NIC (not supported yet) IDD Guest Domain I/O Channel eth0 Bridge Vif0 Backend VCPU0 Vif frontend Interrupts Guest Domain I/O Channel eth1 Vif1 Backend VCPU1 Vif frontend

9. 9 Factoring Interference among VMs Cache interference among VMs can degrade performance significantly −Configurations must factor potential interference among VMs Example: Multiple VMs, each compiling the Linux kernel −Comparison against equivalent Linux and single VM configurations Execution Time Comparison Cache and TLB miss counts per VM

10. 10 Summary and Future Work Summary Findings provide guidance to configure Xen: −CPU requirement for I/O processing can be estimated from application traffic profile −I/O processing is limited by the capacity of a single CPU −Must account for degradation due to the sharing of resources that are not explicitly managed by the VMM Exposed scalability and efficiency problems in Xen’s I/O model – needs improvement Future work Extend study to workloads and platforms of typical customer environments (e.g., HP-IT) Develop an automated configuration tool