1 / 23

SUM305 XenServer performance, reality and myths

SUM305 XenServer performance, reality and myths. Radoslaw Smigielski XenServer/UNIX Escalation Engineer 2011.10.25. Agenda. Scalability limits Network performance Dom0 memory adjustments ?????. Citrix Confidential - Do Not Distribute. Scalability Limits – facts and reality.

kiri
Download Presentation

SUM305 XenServer performance, reality and myths

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. SUM305 XenServer performance, reality and myths Radoslaw SmigielskiXenServer/UNIX Escalation Engineer2011.10.25

  2. Agenda Scalability limits Network performance Dom0 memory adjustments ????? Citrix Confidential - Do Not Distribute

  3. Scalability Limits – facts and reality

  4. Scalability Limits - facts Consolidation race and new records every month Impressive results, 16 vCPU on single pCPU Quote “On average they are running 170 VMs per host.(...)” Q: Can I achieve the same results in my environment? There are too many variables in this equation Citrix Confidential - Do Not Distribute

  5. Scalability Limits - facts How to read and understand all those results? What do they really represent? What is the purpose of these benchmarks? Which hypervisors were used to obtain these results and on what hardware? Is it VM per pCPU or vCPU per pCPU? Is it a vanilla system or tweaked? Citrix Confidential - Do Not Distribute

  6. Scalability Limits - facts Experiment, CPU usage of light user Browser up to 14% VLC player up to 2% Power Point up to 6.5% Outlook between 0.2 – 15% iTunes between 0.1 – 7% Acrobat Reader between 0.5 – 8% Java between 0.1 – 60% Antivirus between 1 – 18% • Single CPU utilization in 1 second time interval • BUSY 230 ms • IDLE 770 ms Citrix Confidential - Do Not Distribute

  7. Scalability Limits - facts Math • Light user on single CPU VM • Estimated number of VM or vCPU per pCPU • Know your users’ workload Citrix Confidential - Do Not Distribute

  8. Scalability Limits - facts vCPU density Citrix Confidential - Do Not Distribute

  9. Scalability Limits - myths There is no universal recipe or generic workload Citrix Confidential - Do Not Distribute

  10. Xen scheduler - facts Most popular questions Scheduler distinquishes between physical cores and hyperthreaded cores Algorithm gives preference to hyperthreaded core if second core is idle Migrate only if the other core is twice as idle Citrix Confidential - Do Not Distribute

  11. Network performance

  12. Network stack design - fact Network flow diagram Citrix Confidential - Do Not Distribute

  13. Network stack design - facts So where are the bottlenecks? Fixed number of netback drivers in dom0 Static VIF allocation Hardware IRQ handling in dom0 Citrix Confidential - Do Not Distribute

  14. Network stack design - facts Fixed number of netback drivers in dom0 Citrix Confidential - Do Not Distribute

  15. Network stack design - facts Static VIF allocation Control domain – dom0 VM1 netback0 VM2 netback1 VM3 Netback2 VM4 netback3 Citrix Confidential - Do Not Distribute

  16. Network stack design - facts Hardware IRQ handling in dom0, irqbalance. XenServer 6.0 XenServer 5.6.X Citrix Confidential - Do Not Distribute

  17. Jumbo (Frame) – myths and facts Myth 1 – Always speed up your network performance Myth 2 – Decrease CPU usage Fact 1 – Improve processing big packets, eg storage traffic, NFS (8,192-byte blocks !!!) Fact 2 – Doesn’t improve network speed always, eg ICA Citrix Confidential - Do Not Distribute

  18. Dom0 memory adjustments Dom0 is 32 bit Standard 32 bit systems can address up to 4GB of RAM PAE is enabled in Dom0 PAE makes your page tables larger but slows the system down General recommendation is that you don't use more than 8GiB on a 32-bit (Linux kernel docs highmem.txt) Citrix Confidential - Do Not Distribute

  19. VHD speed Facts and myths VHD is slower than LVM Our tests indicate read/write speed is the same for raw LUN and VHD (zero chain) Performance degradation starts when chain length > 1 Hard-coded limit on length of VHD chain, 30 Citrix Confidential - Do Not Distribute

  20. CPU hyperthreading Facts and reality Duplicates architectural state (general purpose registers, MMU, Interrupt Mask) Does not duplicate execution resources, single execution contexts Intel claims up to 30% performance improvement AX,BX,CX,DX,etc EFLAGS Interrupt Mask MMU Status Registers AX,BX,CX,DX,etc EFLAGS Interrupt Mask MMU Status Registers Citrix Confidential - Do Not Distribute

  21. CPU hyperthreading Dual core vs hyperthreading? Citrix Confidential - Do Not Distribute

  22. SSD support XenServer does support SSD drives connected on SATA, SAS or SCSI bus SSD on PCI-Express cards won’t work TRIM not supported yet Citrix Confidential - Do Not Distribute

More Related