1. Measuring PROOF Lite performance in (non)virtualized environment Ioannis Charalampidis, Aristotle University of Thessaloniki Summer Student 2010

2. Overview Introduction Benchmarks: Overall execution time Benchmarks: In-depth analysis Conclusion

3. What am I looking for? There is a known overhead caused by the virtualization process ▫ How big is it? ▫ Where is located? ▫ How can we minimize it? ▫ Which hypervisor has the best performance? I am using CernVM as guest

4. What is CernVM? It’s a baseline Virtual Software Appliance for use by LHC experiments It’s available for many hypervisors Hyper-V VM Ware Virtual Box KVM / QEMU XEN

5. How am I going to find the answers? Using as benchmark a standard data analysis application (ROOT + PROOF Lite) Test it on different hypervisors And on varying number of workers/CPUs Compare the performance (Physical vs. Virtualized)

6. Problem The benchmark application requires too much time to complete ( 2 min ~ 15 min ) ▫ At least 3 runs are required for reliable results ▫ The in-depth analysis overhead is about 40% ▫ It is not efficient to perform detailed analysis for every CPU / Hypervisor configuration  Create the overall execution time benchmarks  Find the best configuration to run the traces on

7. Benchmarks performed Overall time ▫ Using time utility and automated batch scripts In-depth analysis ▫ Tracing system calls using  Strace  KernelTAP ▫ Analyzing the trace files using applications I wrote  BASST (Batch analyzer based on STrace)  KARBON (General purpose application profiler based on trace files)

8. Process description and results

9. Benchmark Configuration Base machine ▫ Scientific Linux CERN 5 Guests ▫ CernVM 2.1 Software packages from SLC repositories ▫ Linux Kernel 2.6.18-194.8.1.el5 ▫ XEN 3.1.2 + 2.6.18-194.8.1.el5 ▫ KVM 83-194.8.1.el5 ▫ Python 2.5.4p2 (from AFS) ▫ ROOT 5.26.00b (from AFS) Base machine hardware ▫ 24 x Intel Xeon X7460 2.66GHz with VT-x Support (64 bit) ▫ No VT-d nor Extended Page Tables (EPT) hardware support ▫ 32G RAM

10. Benchmark Configuration Virtual machine configuration ▫ 1, 2 to 16 CPUs with 2 CPU step ▫ + 1Gb RAM for Physical disk and Network tests ▫ + 17Gb RAM for RAM Disk tests ▫ Disk image for the OS ▫ Physical disk for the Data + Software Important background services running ▫ NSCD (Caching daemon)

11. Benchmark Configuration Caches were cleared before every test ▫ Page cache, dentries and inodes ▫ Using the /proc/sys/vm/drop_caches flag No swap memory was used ▫ By periodically monitoring the free memory

12. Automated batch scripts The VM batch script runs on the host machine It repeats the following procedure: ▫ Crate a new Virtual Machine ▫ Wait for the machine to finish booting ▫ Connect to the controlling script inside the VM ▫ Drop caches both on the host and the guest ▫ Start the job ▫ Receive and archive the results Client Server Hypervisor Benchmark

13. Problem There was a bug on PROOF Lite that was looking up a non-existing hostname during the startup of each worker  Example : 0.2-plitehp24.cern.ch-1281241251-1271 Discovered by detailed system call tracing  The hostname couldn’t be cached  The application had to wait for the timeout  The startup time was delayed randomly  Call tracing applications made this delay even bigger virtually hanging the application

14. Problem The problem was resolved with: ▫ A minimal DNS proxy was developed that fakes the existence of the buggy hostname ▫ It was later fixed in PROOF source Application DNS Server Fake DNS Proxy cernvm.cern.ch? 137.138.234.20 x.x-xxxxxx-xxx-xxx? 127.0.0.1

15. Problem Example: Events / sec for different CPU settings, as reported by the buggy benchmark BeforeAfter

16. Results – Physical Disk

17. Results – Network (XROOTD)

18. Results – RAM Disk

19. Results – Relative values RAM DiskNetwork (XROOTD) Physical Disk Bare metal KVMXEN

20. Results – Absolute values RAM DiskNetwork (XROOTD) Physical Disk Bare metal KVMXEN

21. Results – Comparison chart

22. Procedure, problems and results

23. In depth analysis In order to get more details the program execution was monitored and all the system calls were traced and logged Afterwards, the analyzer extracted useful information from the trace files such as ▫ Detecting the time spent on each system call ▫ Detecting the filesystem / network activity The process of tracing adds some overhead but it is cancelled out from the overall performance measurement

24. System call tracing utilities STrace ▫ Traces application-wide system calls from user space ▫ Connects to the tracing process using the ptrace() system call and monitors it’s activity Advantages ▫ Traces the application’s system calls in real time ▫ Has very verbose output Disadvantages ▫ Creates big overhead Process Kernel STrace

25. System call tracing utilities SystemTAP ▫ Traces system-wide kernel activity, asynchronously ▫ Runs as a kernel module Advantages ▫ Can trace virtually everything on a running kernel ▫ Supports scriptable kernel probes Disadvantages ▫ It is not simple to extract detailed information ▫ System calls can be lost on high CPU activity Process Kernel System TAP

26. System call tracing utilities Sample STrace output: 5266 1282662179.860933 arch_prctl(ARCH_SET_FS, 0x2b5f2bcc27d0) = 0 5266 1282662179.860960 mprotect(0x34ca54d000, 16384, PROT_READ) = 0 5266 1282662179.860985 mprotect(0x34ca01b000, 4096, PROT_READ) = 0 5266 1282662179.861009 munmap(0x2b5f2bc92000, 189020) = 0 5266 1282662179.861082 open("/usr/lib/locale/locale-archive", O_RDONLY) = 4 5266 1282662179.861113 fstat(4, {st_mode=S_IFREG|0644, st_size=56442560,...}) = 0 5266 1282662179.861166 mmap(NULL, 56442560, PROT_READ, MAP_PRIVATE, 4, 0) = 0x2b5f2bcc3000 5266 1282662179.861192 close(4) = 0 5266 1282662179.861269 brk(0) = 0x1ad1f000 5266 1282662179.861290 brk(0x1ad40000) = 0x1ad40000 5266 1282662179.861444 open("/usr/share/locale/locale.alias", O_RDONLY) = 4 5266 1282662179.861483 fstat(4, {st_mode=S_IFREG|0644, st_size=2528,...}) = 0 5266 1282662179.861944 read(4, "", 4096) = 0 5266 1282662179.861968 close(4) = 0 5266 1282662179.861989 munmap(0x2b5f2f297000, 4096) = 0 5264 1282662179.863063 wait4(-1, 0x7fff8d813064, WNOHANG, NULL) = -1 ECHILD (No child processes)...

27. KARBON – A trace file analyzer

28. Is a general purpose application profiler based on system call trace files It traces file descriptors and reports detailed I/O statistics for files, network sockets and FIFO pipes It analyzes the child processes and creates process graphs and process trees It can detect the “Hot spots” of an application Custom analyzing tools can be created on-demand using the development API

29. KARBON – Application block diagram Source (File or TCP Stream) Router Preprocessing Tool Analyzer Filter Presenter Tokenizer

30. Results Time utilization of the traced application

31. Results Time utilization of the traced application

32. Results Time utilization of the traced application

33. Results Overall system call time for filesystem I/O Reminder: Kernel buffers were dropped before every test ▫ Possible caching effect inside the hypervisor [ms]ReadingWritingSeekingTotal Bare metal490,861.3542,054.35421,594.583524,872.823 KVM38,391.71536,422.440122,769.518244,406.512 XEN38,111.98020,930.382102,769.901210,247.468

34. Results Overall system call time for UNIX Sockets [ms]ReceivingSendingBind, ListenConnectingTotal Bare metal993.88410,313.3044.2515.25911,301.588 KVM59,637.942164,655.0777.41213.656223,872.164 XEN97,823.986550,050.4845.0148.493652,784.010

35. Results Most time-consuming miscellaneous system calls System callBare metalKVMXEN wait4() 178,200.34316,829,30388,885,57 gettimeofday()(No trace) 219,780,33218,018,63 nanosleep()(No trace) 12,250,1212,029,30 time()(No trace) 9,081,94 rt_sigreturn() 150,9431,685,2859,271,061 setitimer() 23,245698,785223,669

36. Conclusion Physical Disk ▫ KVM can achieve better performance than XEN, reaching 70 - 98% of the native speed ▫ Best performance achieved on 6 CPUs/6 workers (7Gb RAM) with 81% of the native speed Network (Xrootd) ▫ XEN can achieve better performance than KVM, reaching 73 - 90% of the native speed ▫ Best performance achieved again on 6 CPUs / 6 workers (7G RAM) with 92% of the native speed

37. Conclusion Some disk I/O operations (read) appear to be faster inside the Virtual Machine Some of them appear to be slower (seek, write) ▫ Possible caching effect even on direct disk access Network I/O ▫ TCP under XEN looks fine, whereas with KVM there are some issues ▫ UNIX Sockets seem to have significant penalty inside the VMs Some miscellaneous system calls take longer inside the VM ▫ Time-related functions (gettimeoftheday, nanosleep)  Used for paravirtualized implementation of other system calls?

38. Other uses of the tools SystemTAP could be used by nightly builds in order to detect hanged applications KARBON can be used as a general log file analysis program

39. Future work Benchmark VMs with a disk image file residing on a RAID Array Benchmark many concurrent KVM virtual machines with total memory exceed the overall system memory – Exploit NPT Test the PCI Pass-through for network cards (KVM) – Test VT-d Convert the benchmark application from python to pure C Repeat the benchmarks with the optimized ROOT input files Test again the KVM Network performance with Recompile the kernel with CONFIG_KVM_CLOCK