SCaLE12x (2014): What Linux can learn from Solaris performance and vice-versa
Keynote for the Southern California Linux Expo 12x (2014) by Brendan Gregg.Video: http://www.youtube.com/watch?v=6TYC5h4yz1o
What Linux can learn from Solaris performance and vice-versa.
PDF: SCaLE_Linux_vs_Solaris_Performance2014.pdf
Keywords (from pdftotext):
slide 1:
What Linux can learn from Solaris performance and vice-versa Brendan Gregg Lead Performance Engineer brendan@joyent.com @brendangregg SCaLE12x February, 2014slide 2:
Linux vs Solaris Performance Differences DTrace libumem Symbols Microstate futex Accounting Up-to-date Mature fully packages preemptive Applications CPU scalability likely()/unlikely() CONFIGurable RCU DynTicks System Libraries SLUB System Call Interface ZFS btrfs I/O Scheduler VFS Sockets File Systems TCP/UDP Volume Managers Block Device Interface Ethernet Scheduler Virtual Memory Virtualization Device Drivers Zones KVM Crossbow FireEngine More device drivers Process swapping Overcommit & OOM Killer MPSS Lazy TLBslide 3:
whoami • Lead Performance Engineer at Joyent • Work on Linux and SmartOS performance • Work/Research: tools, visualizations, methodologies • Did kernel engineering at Sun Microsystems; worked on DTrace and ZFSslide 4:
Joyent • High-Performance Cloud Infrastructure • Compete on cloud instance/OS performance • OS Virtualization for bare metal performance (Zones) • Core developers of illumos/SmartOS and Node.js • Recently launched Manta: a high performance object store • KVM for Linux guests • Certified Ubuntu on Joyent cloud now available!slide 5:
SCaLE11x: Linux Performance Analysis strace Operating System pidstat netstat Hardware perf Applications DBs, all server types, ... mpstat System Libraries CPU Interconnect System Call Interface Linux Kernel perf dtrace stap lttng ktap VFS Sockets File Systems TCP/UDP Volume Managers Block Device Interface Ethernet Scheduler top ps pidstat Virtual Memory vmstat slabtop free Device Drivers iostat iotop blktrace perf Expander Interconnect I/O Bus I/O Bridge tcpdump I/O Controller Disk Memory Bus perf DRAM nicstat Network Controller Interface Transports Disk CPU Various: Port Swap swapon ping Latest version: http://www.brendangregg.com/linuxperf.html Port traceroute sar dstat /procslide 6:
SCaLE12x: Linux and Solaris Performance • Also covered in my new book, Systems Performance (Prentice Hall, 2013) • Focus is on understanding systems and the methodologies to analyze them. Linux and Solaris are used as examplesslide 7:
Agenda • Why systems differ • Specific differences • What Solaris could learn from Linux • What Linux could learn from Solaris • What both can learn • Resultsslide 8:
Terminology • For convenience in this talk: • Linux = an operating system distribution which uses the Linux kernel. eg, Ubuntu. • Solaris = a distribution from the family of operating systems whose kernel code originated from Sun Solaris. • SmartOS = a Solaris-family distribution developed by Joyent, based on the illumos kernel, which was based on the OpenSolaris kernel, which was based on the Solaris kernel • System = everything you get when you pick a Linux or Solaris distribution: the kernel, libraries, tools, and package repos • Opinions in this presentation are my own, and I do not represent Oracle Solaris. I'll actually be talking about SmartOS.slide 9:
Why Systems Differslide 10:
Why Systems Differ • Does the system even matter? • Will your application perform the same on Linux and Solaris?slide 11:
Example • Let's start with this simple program: perl -e 'for ($i = 0; $islide 12: Example Results • One of these is Linux, the other SmartOS. Same hardware: systemA$ time perl -e 'for ($i = 0; $islide 13: Possible Differences:Versioning • Different versions of Perl • Applications improve performance from release to release • Linux and SmartOS distributions use entirely different package repos; different software versions are common • Different compilers used to build Perl • Compilers come from package repos, too. I've seen 50% performance improvements by gcc version alone • Different compiler options used to build Perl • Application Makefile: #ifdef Linux -O3 #else -O0. ie, the performance difference is due to a Makefile decision • 32-bit vs 64-bit buildsslide 14:Possible Differences: OS • Different system libraries • If any library calls are used. eg: strcmp(), malloc(), memcpy(), ... These implementations vary between Linux and Solaris, and can perform very differently • Robert Mustacchi enhanced libumem to provide improved malloc() performance on SmartOS. This can make a noticeable difference for some workloads • Different background tasks • Runtime could be perturbed by OS daemons doing async housekeeping. These differ between Linux and Solarisslide 15:Possible Differences: Observability • Can the 14% be root caused? • Observability tools differ. These don't cause the 14%, but can make the difference as to whether you can diagnose and fix it – or not. • DTrace has meant that anything can be solved; without an equivalent on Linux, you may have to live with that 14% • Although, Linux observability is getting much better...slide 16:Possible Differences: Kernel • Can the kernel make a difference? ... As a reminder: perl -e 'for ($i = 0; $islide 17: Possible Differences: Kernel, cont. • Can the kernel make a difference? ... As a reminder: perl -e 'for ($i = 0; $islide 18: Possible Differences: Kernel, cont. • During a perturbation, the kernel CPU scheduler may migrate the thread to another CPU, which can hurt performance (cold caches, memory locality) • Sure, but would that happen for this simple Perl program?slide 19:Possible Differences: Kernel, cont. • During a perturbation, the kernel CPU scheduler may migrate the thread to another CPU, which can hurt performance (cold caches, memory locality) • Sure, but would that happen for this simple Perl program? # dtrace -n 'profile-99 /pid == $target/ { @ = lquantize(cpu, 0, 16, 1); }' -c ... value ------------- Distribution ------------- countslide 20: Kernel Myths and Realities • Myth: "The kernel gets out of the way for applications" • The only case where the kernel gets out of the way is when your software calls halt() or shutdown() • The performance difference between kernels may be small, eg, 5% – but I have already seen a 5x difference this year arch/ia64/kernel/smp.c: void cpu_die(void) max_xtp(); local_irq_disable(); cpu_halt(); /* Should never be here */ BUG(); for (;;); unintentional kernel humor...slide 21:Other Differences • The previous example was simple. Any applications that do I/O (ie, everything) encounter more differences: • Different network stack implementations, including support for different TCP/IP features • Different file systems, storage I/O stacks • Different device drivers and device feature support • Different resource control implementations • Different virtualization technologies • Different community support: stackoverflow, meetups, ...slide 22:Types of Differences App versions from system repos Syscall interface File systems: ZFS, btrfs, ... I/O scheduling Virtualization Technologies Resource controls Compiler Observability System library options tools daemons implementations; malloc(), str*(), ... Applications DBs, all server types, ... System Libraries System Call Interface VFS Sockets File Systems TCP/UDP Volume Managers Block Device Interface Ethernet Scheduler Virtual Memory Virtualization Device Drivers Device driver TCP/IP stack support and features Scheduler classes and behavior Memory allocation and locality Network device CPU fanoutslide 23:Specific Differencesslide 24:Specific Differences • Comparing systems is like comparing countries • I'm often asked: how's Australia different from the US? • Where do I start!? • I'll categorize performance differences into big or small, based on their engineering cost, not their performance effect • If one system is 2x faster than another for a given workload, the real question for the slower system is: • Is this a major undertaking to fix? • Is there a quick fix or workaround? • Using SmartOS for specific examples...slide 25:Big Differences • Major bodies of perf work and other big differences, include: • Linux • up-to-date packages, large community, more device drivers, futex, RCU, btrfs, DynTicks, SLUB, I/O scheduling classes, overcommit & OOM killer, lazy TLB, likely()/ unlikely(), CONFIGurable • SmartOS • Mature: Zones, ZFS, DTrace, fully pre-emptable kernel • Microstate accounting, symbols by default, CPU scalability, MPSS, libumem, FireEngine, Crossbow, binary /proc, process swappingslide 26:Big Differences: Linux Up-to-date packages Large community More device drivers Latest application versions, with the latest performance fixes Weird perf issue? May be answered on stackoverflow, or discussed at meetups There can be better coverage for high performing network cards or driver features futex Fast user-space mutex RCU Fast-performing read-copy updates btrfs Modern file system with pooled storage DynTicks SLUB Dynamic ticks: tickless kernel, reduces interrupts and saves power Simplified version of SLAB kernel memory allocator, improving performanceslide 27:Big Differences: Linux, cont. I/O scheduling classes Block I/O classes: deadline, anticipatory, ... Overcommit & OOM killer Doing more with less main memory Lazy TLB Higher performing munmap() likely()/unlikely() CONFIGurable Kernel is embedded with compiler information for branch prediction, improving runtime perf Lightweight kernels possible by disabling featuresslide 28:Big Differences: SmartOS OS virtualization for high-performing server Mature Zones instances Fully-featured and high-performing modern Mature ZFS integrated file system with pooled storage Programmable dynamic and static tracing for Mature DTrace performance analysis Mature fully preSupport for real-time systems was an early Sun emptable kernel differentiator Numerous high-resolution thread state times for Microstate accounting performance debugging Symbols Symbols available for profiling tools by default CPU scalability Code is often tested, and bugs fixed, for large SMP servers (mainframes) MPSS Multiple page size support (not just hugepages)slide 29:Big Differences: SmartOS, cont. libumem FireEngine Crossbow binary /proc Process swapping High-performing memory allocation library, with per-thread CPU caches High-performing TCP/IP stack enhancements, including vertical perimeters and IP fanout High-performing virtualized network interfaces, as used by OS virtualization Process statistics are binary (slightly more efficient) by default Apart from paging (what Linux calls swapping), Solaris can still swap out entire processesslide 30:Big Differences: Linux vs SmartOS DTrace libumem Symbols Microstate futex Accounting Up-to-date Mature fully packages preemptive Applications CPU scalability likely()/unlikely() CONFIGurable RCU DynTicks System Libraries SLUB System Call Interface ZFS btrfs I/O Scheduler VFS Sockets File Systems TCP/UDP Volume Managers Block Device Interface Ethernet Scheduler Virtual Memory Resource Controls Device Drivers Zones Crossbow FireEngine More device drivers Process swapping Overcommit & OOM Killer MPSS Lazy TLBslide 31:Small Differences • Smaller performance-related differences, tunables, bugs • Linux • glibc, better TCP defaults, better CPU affinity, perf stat, a working sar, htop, splice(), fadvise(), ionice, /usr/bin/time, mpstat %steal, voluntary preemption, swappiness, various accounting frameworks, tcp_tw_reuse/recycle, TCP tail loss probe, SO_REUSEPORT, ... • SmartOS • perf tools by default, kstat, vfsstat, iostat -e, ptime -m, CPU-only load averages, some STREAMS leftovers, ZFS SCSI cache flush by default, different TCP slow start default, ...slide 32:Small Differences, cont. • Small differences change frequently: a feature is added to one kernel, then the other a year later; a difference may only exist for a short period of time. • These small kernel differences may still make a significant performance difference, but are classified as "small" based on engineering cost.slide 33:System Similarities • It's important to note that many performance-related features are roughly equivalent: • Both are Unix-like systems: processes, kernel, syscalls, time sharing, preemption, virtual memory, paged virtual memory, demand paging, ... • Similar modern features: unified buffer cache, memory mapped files, multiprocessor support, CPU scheduling classes, CPU sets, 64-bit support, memory locality, resource controls, PIC profiler, epoll, ...slide 34:Non Performance Differences • Linux • Open source (vs Oracle Solaris), "everyone knows it", embedded Linux, popular and well supported desktop/ laptop use... • SmartOS • SMF/FMA, contracts, privileges, mdb (postmortem debugging), gcore, crash dumps by default, ...slide 35:WARNING The next sections are not suitable for those suffering Not Invented Here (NIH) syndrome, or those who are easily trolledslide 36:What Solaris can learn from Linux performanceslide 37:What Solaris can learn from Linux performance • Packaging • Overcommit & OOM Killer • Community • SLUB • Compiler Options • Lazy TLB • likely()/unlikely() • TIME_WAIT Recycling • Tickless Kernel • sar • Process Swapping • KVM • Either learning what to do, or learning what not to do...slide 38:Packaging • Linux package repositories are often well stocked and updated • Convenience aside, this can mean that users run newer software versions, along with the latest perf fixes • They find "Linux is faster", but the real difference is the version of: gcc, openssl, mysql, ... Solaris is unfairly blamedslide 39:Packaging, cont. • Packaging is important and needs serious support • Dedicated staff, community • eg, Joyent has dedicated staff for the SmartOS package repo, which is based on pkgsrc from NetBSD • It's not just the operating system that matters; it's the ecosystemslide 40:Community • A large community means: • Q/A sites have performance tips: stackoverflow, ... • Conference talks on performance (this one!), slides, video • Weird issues more likely found and fixed by someone else • More case studies shared: what tuning/config worked • A community helps people hear about the latest tools, tuning, and developments, and adopt themslide 41:Community, cont. • Linux users expect to Google a question and find an answer on stackoverflow • Either foster a community to share content on tuning, tools, configuration, or, have staff to create such content. • Hire a good community manager!slide 42:Compiler Options • Apps may compile with optimizations for Linux only. eg: Oh, ha ha ha • #ifdef Linux -O3 #else -O0 • Developers are often writing software on Linux, and that platform gets the most attention. (Works on my system.) • I've also seen 64-bit vs 32-bit. #ifdef Linux USE_FUTEX would be fine, since Solaris doesn't have them yet. • Last time I found compiler differences using Flame Graphs: Extra Function: UnzipDocid() Linux SmartOSslide 43:Compiler Options, cont. • Can be addressed by tuning packages in the repo • Also file bugs/patches with developers to tune Makefiles • Someone has to do this, eg, package repo staff/community who find and do the workarounds anywayslide 44:likely()/unlikely() • These become compiler hints (__builtin_expect) for branch prediction, and are throughout the Linux kernel: net/ipv4/tcp_output.c, tcp_transmit_skb(): [...] [...] if (likely(clone_it)) { if (unlikely(skb_cloned(skb))) skb = pskb_copy(skb, gfp_mask); else skb = skb_clone(skb, gfp_mask); if (unlikely(!skb)) return -ENOBUFS; • The Solaris kernel doesn't do this yet • If the kernel is built using profile feedback instead – which should be even better – I don't know about it • The actual perf difference is likely to be smallslide 45:likely()/unlikely(), cont. • Could be adopted by kernel engineering • Might help readability, might notslide 46:Tickless Kernel • Linux does this already (DynTicks), which reduces interrupts and improves processor power saving (good for laptops and embedded devices) • Solaris still has a clock() routine, to perform various kernel housekeeping functions • Called by default at 100 Hertz • If hires_tick=1, at 1000 Hertz • I've occasionally encountered perf issues involving 10 ms latencies, that don't exist on Linux • ... which become 1 ms latencies after setting hires_tick=1slide 47:Tickless Kernel, cont. • Sun/Oracle did start work on this years ago...slide 48:Process Swapping • Linux doesn't do it. Linux "swapping" means paging. • Process swapping made sense on the PDP-11/20, where the maximum process size was 64 Kbytes • Paging was added later in BSD, but the swapping code remainedslide 49:Process Swapping, cont. • Consider ditching it • All that time learning what swapping is could be spent learning more useful featuresslide 50:Overcommit & OOM Killer • On Linux, malloc() may never fail • No virtual memory limit (main memory + swap) like Solaris by default. Tunable using vm.overcommit_memory • More user memory can be allocated than can be stored. May be great for small devices, running applications that sparsely use the memory they allocate • Don't worry, if Linux runs very low on available main memory, a sacrificial process is identified by the kernel and killed by the Out Of Memory (OOM) Killer, based on an OOM score • OOM score was just added to htop (1.0.2, Jan 2014):slide 51:Overcommit & OOM Killer, cont. • Solaris can learn why not to do this (cautionary tale) • If an important app depended on this, and couldn't be fixed, the kernel could have an overcommit option that wasn't default • ... this is why so much new code doesn't check for ENOMEMslide 52:SLUB • Linux integrated the Solaris kernel SLAB allocator, then later simplified it: The SLUB allocator • Removed object queues and per-CPU caches, leaving NUMA optimization to the page allocator's free lists • Worth considering?slide 53:Lazy TLB • Lazy TLB mode: a way to delay TLB updates (shootdowns) • munmap() heavy workloads on Solaris can experience heavy HAT CPU cross calls. Linux doesn't seem to have this problem. TLB Lazy TLB As seen by Solaris Correct Reckless As seen by Linux Paranoid Fastslide 54:Lazy TLB, cont. • This difference needs to be investigated, quantified, and possibly fixed (tunable?)slide 55:TIME_WAIT Recycling • A localhost HTTP benchmark on Solaris: # netstat -s 1 | grep ActiveOpen tcpActiveOpens =728004 tcpActiveOpens tcpActiveOpens = 4939 tcpActiveOpens = 5849 tcpActiveOpens = 1341 tcpActiveOpens = 1006 tcpActiveOpens tcpActiveOpens tcpActiveOpens tcpActiveOpens tcpActiveOpens tcpPassiveOpens tcpPassiveOpens tcpPassiveOpens tcpPassiveOpens tcpPassiveOpens tcpPassiveOpens tcpPassiveOpens tcpPassiveOpens tcpPassiveOpens tcpPassiveOpens tcpPassiveOpens =726547 = 4939 = 5798 = 1292 = 1008 Fast Slow • Connection rate drops by 5x due to sessions in TIME_WAIT • Linux avoids this by recycling better (tcp_tw_reuse/recycle) • Usually doesn't hurt production workloads, as it must be a flood of connections from a single host to a single port. It comes up in benchmarks/evaluations.slide 56:TIME_WAIT Recycling, cont. • Improve tcp_time_wait_processing() • This is being fixed for illumos/SmartOSslide 57:sar • Linux sar is awesome, and has extra options: $ sar -n DEV -n TCP -n ETCP 1 11:16:34 PM IFACE rxpck/s 11:16:35 PM eth0 11:16:35 PM eth1 11:16:35 PM ip6tnl0 11:16:35 PM 11:16:35 PM ip_vti0 11:16:35 PM sit0 11:16:35 PM tunl0 11:16:34 PM 11:16:35 PM active/s passive/s 11:16:34 PM 11:16:35 PM atmptf/s txpck/s rxkB/s iseg/s oseg/s estres/s retrans/s isegerr/s txkB/s rxcmp/s orsts/s • -n DEV: network interface statistics • -n TCP: TCP statistics • -n ETCP: TCP error statistics • Linux sar's other metrics are also far less buggy txcmp/s rxmcst/sslide 58:sar, cont. • Sar must be fixed for the 21st century • Use the Linux sar options and column names, which follow a neat conventionslide 59:KVM • The KVM type 2 hypervisor originated for Linux • While Zones are faster, KVM can run different kernels (Linux) • vs Type 1 hypervisors (Xen): • KVM has better perf observability, as it can use the regular OS tools • KVM can use OS resource controls, just like any other processslide 60:KVM, cont. • illumos/SmartOS learned this, Joyent ported KVM! • Oracle Solaris doesn't have it yetslide 61:What Linux can learn from Solaris performanceslide 62:What Linux can learn from Solaris performance • ZFS • Zones • STREAMS • Symbols • prstat -mLc • vfsstat • DTrace • Culture • Either learning what to do, or learning what not to do...slide 63:ZFS • More performance features than you can shake a stick at: • Pooled storage, COW, logging (batching writes), ARC, variable block sizes, dynamic striping, intelligent prefetch, multiple prefetch streams, snapshots, ZIO pipeline, compression (lzjb can improve perf by reducing I/O load), SLOG, L2ARC, vdev cache, data deduplication (possibly better cache reach) • The Adaptive Replacement Cache (ARC) can make a big difference: it can resist perturbations (backups) and stay warm • ZFS I/O throttling (in illumos/SmartOS) throttles disk I/O at the VFS layer, to solve cloud noisy neighbor issues • ZFS is Mature. Widespread use in criticial environmentsslide 64:ZFS, cont. • Linux has been learning about ZFS for a while • http://zfsonlinux.org/ • btrfsslide 65:Zones • Ancestry: chroot FreeBSD jails Solaris Zones • OS Virtualization. Zero I/O path overheads.slide 66:Zones, cont. • Compare to HW Virtualization: • This shows the initial I/O control flow. There are optimizations/ variants for improving the HW Virt I/O path, esp for Xen.slide 67:Zones, cont. • Comparing 1 GB instances on Joyent • Max network throughput: • KVM: 400 Mbits/sec • Zones: 4.54 Gbits/sec (over 10x) • Max network IOPS: • KVM: 18,000 packets/sec • Zones: 78,000 packets/sec (over 4x) • Numbers go much higher for larger instances • http://dtrace.org/blogs/brendan/2013/01/11/virtualization-performance-zones-kvm-xenslide 68:Zones, cont. • Performance analysis for Zones is also easy. Analyze the applications as usual: Operating System Applications Zone . analyze System Libraries Kernel System Call Interface VFS Sockets File Systems TCP/UDP Volume Managers Block Device Interface Ethernet Resource Controls Device Drivers Firmware Metal Scheduler Virtual Memoryslide 69:Zones, cont. Host Applications • Compared QEMU to HW Virt (KVM): analyze Guest Applications System Libraries System Call Interface Linux kernel VFS Sockets File Systems TCP/UDP Volume Managers Block Device Interface Ethernet Scheduler Virtual Memory Resource Controls correlate Device Drivers observability boundary Virtual Devices System Libraries System Call Interface host kernel KVM VFS Sockets File Systems TCP/UDP Volume Managers Block Device Interface Ethernet Resource Controls Device Drivers Firmware Metal Scheduler Virtual Memoryslide 70:Zones, cont. • Linux has been learning: LXC & cgroups, but not widespread adoption yet. Docker will likely drive adoption.slide 71:STREAMS • AT&T modular I/O subsystem • Like Unix shell pipes, but for kernel messages. Can push modules into the stream to customize processing • Introduced (fully) in Unix 8th Ed Research Unix, became SVr4 STREAMS, and was used by Solaris for network TCP/IP stack • With greater demands for TCP/IP performance, the overheads of STREAMS reduced scalability • Sun switched high-performing paths to be direct function callsslide 72:STREAMS, cont. • A cautionary tale: not good for high performance code pathsslide 73:Symbols • Compilers on Linux strip symbols by default, making perf profiler output inscrutable without the dbgsym packages 57.14% sshd libc-2.15.so [.] connect --- connect |--25.00%-- 0x7ff3c1cddf29 |--25.00%-- 0x7ff3bfe82761 What?? 0x7ff3bfe82b7c |--25.00%-- 0x7ff3bfe82dfc --25.00%-- [...] • Linux compilers also drop frame pointers, making stacks hard to profile. Please use -fno-omit-frame-pointer to stop this. • as a workaround, perf_events has "-g dwarf" for libunwind • Solaris keeps symbols and stacks, and often has CTF too, making Mean Time To Flame Graph very fastslide 74:Symbols, cont. Flame Graphs need symbols and stacksslide 75:Symbols, cont. • Keep symbols and frame pointers. Faster resolution for performance analysis and troubleshooting.slide 76:prstat -mLc • Per-thread time broken down into states, from a top-like tool: $ prstat -mLc 1 PID USERNAME USR SYS TRP TFL DFL LCK SLP LAT VCX ICX SCL SIG PROCESS/LWPID 63037 root 83 16 0.1 0.0 0.0 0.0 0.0 0.5 30 243 45K 0 node/1 12927 root 14 49 0.0 0.0 0.0 0.0 34 2.9 6K 365 .1M 0 ab/1 63037 root 0.5 0.6 0.0 0.0 0.0 3.7 95 0.4 1K 0 1K 0 node/2 [...] • These columns narrow an investigation immediately, and have been critical for solving countless issues. Unsung hero of Solaris performance analysis • Well suited for the Thread State Analysis (TSA) methodology, which I've taught in class, and has helped students get started and fix unknown perf issues • http://www.brendangregg.com/tsamethod.htmlslide 77:prstat -mLc, cont. • Linux has various thread states: delay accounting, I/O accounting, schedstats. Can they be added to htop? See TSA Method for use case and desired metrics.slide 78:vfsstat • VFS-level iostat (added to SmartOS, not Solaris): $ vfsstat -M 1 r/s w/s Mr/s Mw/s ractv wactv read_t writ_t 761.0 267.1 15.4 4076.8 2796.0 41.7 2.3 0.1 4945.1 2807.4 157.1 2.3 0.1 3550.9 1910.4 109.7 1.6 0.4 0.0 112.9 [...] d/s del_t zone 23.5 5716a5b6 0.0 5716a5b6 0.0 5716a5b6 0.0 5716a5b6 • Shows what the applications request from the file system, and the true performance that they experience • iostat includes asynchronous I/O • vfsstat sees issues iostat can't: • lock contention • resource control throttling Applications System Libraries System Call Interface vfsstat VFS File Systems Volume Managers iostat Block Device Interface Device Drivers Storage Devicesslide 79:vfsstat, cont. • Add vfsstat, or VFS metrics to sar.slide 80:DTrace • Programmable, real-time, dynamic and static tracing, for performance analysis and troubleshooting, in dev and production • Used on Solaris, illumos/SmartOS, Mac OS X, FreeBSD, ... • Solve virtually any perf issue. eg, fix the earlier Perl 15% delta, no matter where the problem is. Without DTrace's capabilities, you may have to wear that 15%. • Users can write their own custom DTrace one-liners and scripts, or use/modify others (eg, mine).slide 81:DTrace: illumos Scripts • Some of my DTrace scripts from the DTraceToolkit, DTrace book... cifs*.d, iscsi*.d :Services nfsv3*.d, nfsv4*.d ssh*.d, httpd*.d Language Providers: Databases: fswho.d, fssnoop.d sollife.d solvfssnoop.d dnlcsnoop.d zfsslower.d ziowait.d ziostacks.d spasync.d metaslab_free.d iosnoop, iotop disklatency.d satacmds.d satalatency.d scsicmds.d scsilatency.d sdretry.d, sdqueue.d ide*.d, mpt*.d hotuser, umutexmax.d, lib*.d node*.d, erlang*.d, j*.d, js*.d php*.d, pl*.d, py*.d, rb*.d, sh*.d mysql*.d, postgres*.d, redis*.d, riak*.d opensnoop, statsnoop errinfo, dtruss, rwtop rwsnoop, mmap.d, kill.d shellsnoop, zonecalls.d weblatency.d, fddist Applications DBs, all server types, ... System Libraries System Call Interface VFS Sockets File Systems TCP/UDP Volume Managers Block Device Interface Ethernet Device Drivers Scheduler priclass.d, pridist.d cv_wakeup_slow.d displat.d, capslat.d Virtual Memory minfbypid.d pgpginbypid.d macops.d, ixgbecheck.d ngesnoop.d, ngelink.d soconnect.d, soaccept.d, soclose.d, socketio.d, so1stbyte.d sotop.d, soerror.d, ipstat.d, ipio.d, ipproto.d, ipfbtsnoop.d ipdropper.d, tcpstat.d, tcpaccept.d, tcpconnect.d, tcpioshort.d tcpio.d, tcpbytes.d, tcpsize.d, tcpnmap.d, tcpconnlat.d, tcp1stbyte.d tcpfbtwatch.d, tcpsnoop.d, tcpconnreqmaxq.d, tcprefused.d tcpretranshosts.d, tcpretranssnoop.d, tcpsackretrans.d, tcpslowstart.d tcptimewait.d, udpstat.d, udpio.d, icmpstat.d, icmpsnoop.dslide 82:DTrace, cont. • What Linux needs to learn about DTrace: Feature #1 is production safety • There should be NO risk of panics or freezes. It should be an everyday tool like top(1). • Related to production safety is the minimization of overheads, which can be done with in-kernel summaries. Some of the Linux tools need to learn how to do this, too, as the overheads of dump & post-analysis can get too high. • Features aren't features if users don't use themslide 83:DTrace, cont. • Linux might get DTrace-like capabilities via: • dtrace4linux • perf_events • ktap • SystemTap • LTTng • The Linux kernel has the necessary frameworks which are sourced by these tools: tracepoints, kprobes, uprobes • ... and another thing Linux can learn: • DTrace has a memorable unofficial mascot (the ponycorn by Deirdré Straughan, using General Zoi's pony creator). She's created some for the Linux tools too...slide 84:dtrace4linux • Two DTrace ports in development for Linux: • 1. dtrace4linux • https://github.com/dtrace4linux/linux • Mostly by Paul Fox • Not safe for production use yet; I've used it to solve issues by first reproducing them in the lab • 2. Oracle Enterprise Linux DTrace • Has been steady progress. Oracle Linux 6.5 featured "full DTrace integration" (Dec 2013)slide 85:dtrace4linux: Example • Tracing ext4 read/write calls with size distributions (bytes): #!/usr/sbin/dtrace -s fbt::vfs_read:entry, fbt::vfs_write:entry /stringof(((struct file *)arg0)->gt;f_path.dentry->gt;d_sb->gt;s_type->gt;name) == "ext4"/ @[execname, probefunc + 4] = quantize(arg2); dtrace:::END printa("\n %s %s (bytes)%@d", @); # ./ext4rwsize.d dtrace: script './ext4rwsize.d' matched 3 probes CPU FUNCTION:NAME :END [...] vi read (bytes) value ------------- Distribution ------------- count 128 | 256 | 512 |@@@@@@@ 1024 |@ 2048 | 4096 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 8192 |slide 86:dtrace4linux: Example • Tracing TCP retransmits (tcpretransmit.d for 3.11.0-17): #!/usr/sbin/dtrace -qs dtrace:::BEGIN { trace("Tracing TCP retransmits... Ctrl-C to end.\n"); } fbt::tcp_retransmit_skb:entry { this->gt;so = (struct sock *)arg0; this->gt;d = (unsigned char *)&this->gt;so->gt;__sk_common; /* 1st is skc_daddr */ printf("%Y: retransmit to %d.%d.%d.%d, by:", walltimestamp, this->gt;d[0], this->gt;d[1], this->gt;d[2], this->gt;d[3]); stack(99); # ./tcpretransmit.d Tracing TCP retransmits... Ctrl-C to end. 1970 Jan 1 12:24:45: retransmit to 50.95.220.155, by: kernel`tcp_retransmit_skb kernel`dtrace_int3_handler+0xcc kernel`dtrace_int3+0x3a kernel`tcp_retransmit_skb+0x1 that kernel`tcp_retransmit_timer+0x276 used to kernel`tcp_write_timer kernel`tcp_write_timer_handler+0xa0 kernel`tcp_write_timer+0x6c work... kernel`call_timer_fn+0x36 kernel`tcp_write_timer kernel`run_timer_softirq+0x1fd kernel`__do_softirq+0xf7 kernel`call_softirq+0x1c [...]slide 87:perf_events • In the Linux tree. perf-tools package. Can do sampling, static and dynamic tracing, with stack traces and local variables • Often involves an enablecollectdumpanalyze cycle • A powerful profiler, loaded with features (eg, libunwind stacks!) • Isn't programmable, and so has limited ability for processing data in-kernel. Does counts. • You can post-process in userland, but the overheads of passing all event data incurs overhead; can be Gbytes of dataslide 88:perf_events: Example • Dynamic tracing of tcp_sendmsg() with size: # perf probe --add 'tcp_sendmsg size' [...] # perf record -e probe:tcp_sendmsg -a ^C[ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.052 MB perf.data (~2252 samples) ] # perf script # ======== # captured on: Fri Jan 31 23:49:55 2014 # hostname : dev1 # os release : 3.13.1-ubuntu-12-opt [...] # ======== sshd sshd sshd sshd sshd sshd [...] 1301 [001] 1301 [001] 2371 [000] 2372 [000] 2372 [001] 2372 [001] 502.424719: probe:tcp_sendmsg: (ffffffff81505d80) size=b0 502.424814: probe:tcp_sendmsg: (ffffffff81505d80) size=40 502.952590: probe:tcp_sendmsg: (ffffffff81505d80) size=27 503.025023: probe:tcp_sendmsg: (ffffffff81505d80) size=3c0 503.203776: probe:tcp_sendmsg: (ffffffff81505d80) size=98 503.281312: probe:tcp_sendmsg: (ffffffff81505d80) size=2d0slide 89:ktap • A new static/dynamic tracing tool for Linux • Lightweight, simple, based on lua. Uses bytecode for programmable and safe tracing • Suitable for use on embedded Linux • http://www.ktap.org • Features are limited (still in development), but I've been impressed so far • In development, so I can't recommend production use yetslide 90:ktap: Example • Summarize read() syscalls by return value (size/err): # ktap -e 's = {}; trace syscalls:sys_exit_read { s[arg2] += 1 } trace_end { histogram(s); }' value ------------- Distribution ------------- count -11 |@@@@@@@@@@@@@@@@@@@@@@@@ 18 |@@@@@@ histogram 72 |@@ 1024 |@ of a key/ 0 | 2 | value table 446 | 515 | 48 | • Write scripts (excerpt from syslatl.kp, highlighting time delta): trace syscalls:sys_exit_* { if (self[tid()] == nil) { return } delta = (gettimeofday_us() - self[tid()]) / (step * 1000) if (delta >gt; max) { max = delta } lats[delta] += 1 self[tid()] = nilslide 91:ktap: Setup • Installing on Ubuntu (~5 minutes): # apt-get install git gcc make # git clone https://github.com/ktap/ktap # cd ktap # make # make install # make load • Example dynamic tracing of tcp_sendmsg() and stacks: # ktap -e 's = ptable(); trace probe:tcp_sendmsg { s[backtrace(12, -1)]slide 92: SystemTap • Sampling, static and dynamic tracing, fully programmable • The most featured of all the tools. Does some things that DTrace can't (eg, loops). • http://sourceware.org/systemtap • Has its own tracing language, which is compiled (gcc) into kernel modules (slow; safe?) • I used it a lot in 2011, and had problems with panics/freezes; never felt safe to run it on my customer's production systems • Needs vmlinux/debuginfoslide 93:SystemTap: Setup • Setting up a SystemTap on ubuntu (2014): # ./opensnoop.stp semantic error: while resolving probe point: identifier 'syscall' at ./ opensnoop.stp:11:7 source: probe syscall.open helpful tips... semantic error: no match Pass 2: analysis failed. [man error::pass2] Tip: /usr/share/doc/systemtap/README.Debian should help you get started. # more /usr/share/doc/systemtap/README.Debian [...] supported yet, see Debian bug #691167). To use systemtap you need to manually install the linux-image-*-dbg and linux-header-* packages that match your running kernel. To simplify this task you can use the stap-prep command. Please always run this before reporting a bug. # stap-prep You need package linux-image-3.11.0-17-generic-dbgsym but it does not seem to be available Ubuntu -dbgsym packages are typically in a separate repository Follow https://wiki.ubuntu.com/DebuggingProgramCrash to add this repositoryslide 94:SystemTap: Setup, cont. • After following ubuntu's DebuggingProgramCrash site: # apt-get install linux-image-3.11.0-17-generic-dbgsym Reading package lists... Done Building dependency tree but my perf issue Reading state information... Done is happening now... The following NEW packages will be installed: linux-image-3.11.0-17-generic-dbgsym 0 upgraded, 1 newly installed, 0 to remove and 0 not upgraded. Need to get 834 MB of archives. After this operation, 2,712 MB of additional disk space will be used. Get:1 http://ddebs.ubuntu.com/ saucy-updates/main linux-image-3.11.0-17generic-dbgsym amd64 3.11.0-17.31 [834 MB] 0% [1 linux-image-3.11.0-17-generic-dbgsym 1,581 kB/834 MB 0%] 215 kB/s 1h 4min 37s • In fairness: • 1. The Red Hat SystemTap developer's primary focus is to get it working on Red Hat (where they say it works fine) • 2. Lack of CTF isn't SystemTap's fault, as said earlierslide 95:SystemTap: Example • opensnoop.stp: #!/usr/bin/stap probe begin printf("\n%6s %6s %16s %s\n", "UID", "PID", "COMM", "PATH"); probe syscall.open printf("%6d %6d %16s %s\n", uid(), pid(), execname(), filename); • Output: # ./opensnoop.stp UID PID 0 11108 0 11108 0 11108 0 11108 0 11108 0 11108 [...] COMM PATH sshdslide 96:gt; sshd gt; sshd /lib/x86_64-linux-gnu/libwrap.so.0 sshd /lib/x86_64-linux-gnu/libpam.so.0 sshd /lib/x86_64-linux-gnu/libselinux.so.1 sshd /usr/lib/x86_64-linux-gnu/libck-connector.so.0 LTTng • Profiling, static and dynamic tracing • Based on Linux Trace Toolkit (LTT), which dabbled with dynamic tracing (DProbes) in 2001 • Involves an enablestartstopview cycle • Designed to be highly efficient • I haven't used it properly yet, so I don't have an informed opinion (sorry LTTng, not your fault)slide 97:LTTng, cont. • Example sequence: # lttng create session1 # lttng enable-event sched_process_exec -k # lttng start # lttng stop # lttng view # lttng destroy session1slide 98:DTrace, cont. • 2014 is an exciting year for dynamic tracing and Linux – one of these may reach maturity and win!slide 99:DTrace, final word • What Oracle Solaris can learn from dtrace4linux: • Dynamic tracing is crippled without source code • Oracle could give customers scripts to run, but customers lose any practical chance of writing them themselves • If the dtrace4linux port is completed, it will be more useful than Oracle Solaris DTrace (unless they open source it again)slide 100:Culture • Sun Microsystems, out of necessity, developed a performance engineering culture that had an appetite for understanding and measuring the system: data-driven analysis • If your several-million-dollar Ultra Enterprise 10000 doesn’t perform well and your company is losing nontrivial sums of money every minute because of it, you call Sun Service and start demanding answers. – System Performance Tuning [Musumeci 02] • Includes the diagnostic cycle: • hypothesis instrumentation data hypothesis • Some areas of Linux are already learning this (and some areas of Solaris never did)slide 101:Culture, cont. • Linux perf issues are often debugged using only top(1), *stat, sar(1), strace(1), and tcpdump(8). These leave many areas not measured. top layer Kernel strace layer If only it were this simple... tcpdump layer • What about the other tools and metrics that are part of Linux? perf_events, tracepoints/kprobes/uprobes, schedstats, I/O accounting, blktrace, etc.slide 102:Culture, cont. • Understand the system, and measure if at all possible • Hypothesis instrumentation data hypothesis • Use perf_events (and others once they are stable/safe) • strace(1) is intermediate, not advanced • High performance doesn't just mean hardware, system, and config. It foremost means analysis of performance limiters.slide 103:What Both can Learnslide 104:What Both can Learn • Get better at benchmarkingslide 105:Benchmarking • How Linux vs Solaris performance is often compared • Results incorrect or misleading almost 100% of the time • Get reliable benchmark results by active benchmarking: • Analyze performance of all components during the benchmark, to identify limiters • Takes time: benchmarks may take weeks, not days • Opposite of passive benchmarking: fire and forget • If SmartOS loses a benchmark, my management demands answers, and I almost always overturn the result with analysis • Test variance as well as steady workloads. Measure jitter. These differ between systems as well.slide 106:Resultsslide 107:Out-of-the-Box • Out-of-the-box, which is faster, Linux or Solaris?slide 108:Out-of-the-Box • Out-of-the-box, which is faster, Linux or Solaris? • There are many differences, it's basically a crapshoot. I've seen everything from 5% to 5x differences either way • It really depends on workload and platform characteristics: • CPU usage, FS I/O, TCP I/O, TCP connect rates, default TCP tunables, synchronous writes, lock calls, library calls, multithreading, multiprocessor, network driver support, ... • From prior Linux vs SmartOS comparisons, it's hard to pick a winner, but in general my expectations are: • SmartOS: for heavy file system or network I/O workloads • Linux: for CPU-bound workloadsslide 109:Out-of-the-Box Relevance • Out-of-the-box performance isn't that interesting: if you care about performance, why not do some analysis and tuning?slide 110:In Practice • With some analysis and tuning, which is faster?slide 111:In Practice • With some analysis and tuning, which is faster? • Depends on the workload, and which differences matter to you • With analysis, I can usually make SmartOS beat Linux • DTrace and microstate accounting give me a big advantage: I can analyze and fix all the small differences (which sometimes exist as apps are developed for Linux) • Although, perf/ktap/... are catching up • I can do the same and make Linux beat SmartOS, but it's much more time-consuming without an equivalent DTrace • On the same hardware, it's more about the performance engineer than the kernel. DTrace doesn't run itself.slide 112:At Joyent • Joyent builds high performance SmartOS instances that frequently beat Linux, but that's due to more than just the OS. We use: • Config: OS virtualization (Zones), all instances on ZFS • Hardware: 10 GbE networks, plenty of DRAM for the ARC • Analysis: DTrace to quickly root-cause issues and tune • With SmartOS (ZFS, DTrace, Zones, KVM) configured for performance, and with some analysis, I expect to win most head-to-head performance comparisonsslide 113:Learning From Linux • Joyent has also been learning from Linux to improve SmartOS. • Package repos: dedicated staff • Community: dedicated staff • Compiler options: dedicated repo staff • KVM: ported! • TCP TIME_WAIT: fixed localhost; more fixes to come • sar: fix understood • More to do...slide 114:References • General Zoi's Awesome Pony Creator: http:// generalzoi.deviantart.com/art/Pony-Creator-v3-397808116 • More perf examples: http://www.brendangregg.com/perf.html • More ktap examples: http://www.brendangregg.com/ktap.html • http://www.ktap.org/doc/tutorial.html • Flame Graphs: http://www.brendangregg.com/flamegraphs.html • Diagnostic cycle: http://dtrace.org/resources/bmc/cec_analytics.pdf • http://www.brendangregg.com/activebenchmarking.html • https://blogs.oracle.com/OTNGarage/entry/doing_more_with_dtrace_onslide 115:Thank You • More info: • illumos: http://illumos.org • SmartOS: http://smartos.org • DTrace: http://dtrace.org • Joyent: http://joyent.com • Systems Performance book: http://www.brendangregg.com/sysperf.html • Me: http://www.brendangregg.com, http://dtrace.org/blogs/ brendan, @brendangregg, brendan@joyent.com