Download as PDF, PPTX
Performance Analysis: The USE Method Brendan Gregg Lead Performance Engineer, Joyent brendan.gregg@joyent.com FISL13 July, 2012
whoami • I work at the top of the performance support chain • I also write open source performance tools out of necessity to solve issues • http://github.com/brendangregg • http://www.brendangregg.com/#software • And books (DTrace, Solaris Performance and Tools) • Was Brendan @ Sun Microsystems, Oracle, now Joyent
Joyent • Cloud computing provider • Cloud computing software • SmartOS • host OS, and guest via OS virtualization • Linux, Windows • guest via KVM
Agenda • Example Problem • Performance Methodology • Problem Statement • The USE Method • Workload Characterization • Drill-Down Analysis • Specific Tools
Example Problem • Recent cloud-based performance issue • Customer problem statement: • "Database response time sometimes take multiple seconds. Is the network dropping packets?" • Tested network using traceroute, which showed some packet drops
Example: Support Path • Performance Analysis Top 2nd Level 1st Level Customer Issues
Example: Support Path • Performance Analysis Top my turn 2nd Level "network looks ok, CPU also ok" 1st Level "ran traceroute, can’t reproduce" Customer: "network drops?"
Example: Network Drops • Old fashioned: network packet capture (sniffing) • Performance overhead during capture (CPU, storage) and post-processing (wireshark) • Time consuming to analyze: not real-time
Example: Network Drops • New: dynamic tracing • Efficient: only drop/retransmit paths traced • Context: kernel state readable • Real-time: analysis and summaries # ./tcplistendrop.d TIME 2012 Jan 19 01:22:49 2012 Jan 19 01:22:49 2012 Jan 19 01:22:49 2012 Jan 19 01:22:49 2012 Jan 19 01:22:49 2012 Jan 19 01:22:49 2012 Jan 19 01:22:49 [...] SRC-IP 10.17.210.103 10.17.210.108 10.17.210.116 10.17.210.117 10.17.210.112 10.17.210.106 10.12.143.16 PORT 25691 18423 38883 10739 27988 28824 65070 -> -> -> -> -> -> -> DST-IP 192.192.240.212 192.192.240.212 192.192.240.212 192.192.240.212 192.192.240.212 192.192.240.212 192.192.240.212 PORT 80 80 80 80 80 80 80
Example: Methodology • Instead of network drop analysis, I began with the USE method to check system health
Example: Methodology • Instead of network drop analysis, I began with the USE method to check system health • In < 5 minutes, I found: • CPU: ok (light usage) • network: ok (light usage) • memory: available memory was exhausted, and the system was paging • disk: periodic bursts of 100% utilization • The method is simple, fast, directs further analysis
Example: Other Methodologies • Customer was surprised (are you sure?) I used latency analysis to confirm. Details (if interesting): • memory: using both microstate accounting and dynamic tracing to confirm that anonymous pagins were hurting the database; worst case app thread spent 97% of time waiting on disk (data faults). • disk: using dynamic tracing to confirm latency at the application / file system interface; included up to 1000ms fsync() calls. • Different methodology, smaller audience (expertise), more time (1 hour).
Example: Summary • What happened: • customer, 1st and 2nd level support spent much time chasing network packet drops. • What could have happened: • customer or 1st level follows the USE method and quickly discover memory and disk issues • memory: fixable by customer reconfig • disk: could go back to 1st or 2nd level support for confirmation • Faster resolution, frees time
Performance Methodology • Not a tool • Not a product • Is a procedure (documentation)
Performance Methodology • Not a tool -> but tools can be written to help • Not a product -> could be in monitoring solutions • Is a procedure (documentation)
Why Now: past • Performance analysis circa ‘90s, metric-orientated: • Vendor creates metrics and performance tools • Users develop methods to interpret metrics • Common method: "Tools Method" • List available performance tools • For each tool, list useful metrics • For each metric, determine interpretation • Problematic: vendors often don’t provide the best metrics; can be blind to issue types
Why Now: changes • Open Source • Dynamic Tracing • See anything, not just what the vendor gave you • Only practical on open source software • Hardest part is knowing what questions to ask
Why Now: present • Performance analysis now (post dynamic tracing), question-orientated: • Users pose questions • Check if vendor has provided metrics • Develop custom metrics using dynamic tracing • Methodologies pose the questions • What would previously be an academic exercise is now practical
Methology Audience • Beginners: provides a starting point • Experts: provides a checklist/reminder
Performance Methodolgies • Suggested order of execution: 1.Problem Statement 2.The USE Method 3.Workload Characterization 4.Drill-Down Analysis (Latency)
Problem Statement • Typical support procedure (1st Methodology): 1.What makes you think there is a problem? 2.Has this system ever performed well? 3.What changed? Software? Hardware? Load? 4.Can the performance degradation be expressed in terms of latency or run time? 5.Does the problem affect other people or applications? 6.What is the environment? What software and hardware is used? Versions? Configuration?
The USE Method • Quick System Health Check (2nd Methodology): • For every resource, check: • Utilization • Saturation • Errors
The USE Method • Quick System Health Check (2nd Methodology): • For every resource, check: • Utilization: time resource was busy, or degree used • Saturation: degree of queued extra work • Errors: any errors Saturation X Errors Utilization
The USE Method: Hardware Resources • CPUs • Main Memory • Network Interfaces • Storage Devices • Controllers • Interconnects
The USE Method: Hardware Resources • A great way to determine resources is to find (or draw) the server functional diagram • The hardware team at vendors should have these • Analyze every component in the data path
The USE Method: Functional Diagrams, Generic Example Memory Bus DRAM CPU Interconnect CPU 1 DRAM CPU 2 I/O Bus I/O Bridge I/O Controller Expander Interconnect Network Controller Interface Transports Disk Disk Port Port
The USE Method: Resource Types • There are two different resource types, each define utilization differently: • I/O Resource: eg, network interface • utilization: time resource was busy. current IOPS / max or current throughput / max can be used in some cases • Capacity Resource: eg, main memory • utilization: space consumed • Storage devices act as both resource types
The USE Method: Software Resources • Mutex Locks • Thread Pools • Process/Thread Capacity • File Descriptor Capacity
The USE Method: Flow Diagram Choose Resource Errors Present? Y N High Utilization? Y N N Saturation? Y Problem Identified
The USE Method: Interpretation • Utilization • 100% usually a bottleneck • 70%+ often a bottleneck for I/O resources, especially when high priority work cannot easily interrupt lower priority work (eg, disks) • Beware of time intervals. 60% utilized over 5 minutes may mean 100% utilized for 3 minutes then idle • Best examined per-device (unbalanced workloads)
The USE Method: Interpretation • Saturation • Any non-zero value adds latency • Errors • Should be obvious
The USE Method: Easy Combinations Resource Type CPU utilization CPU saturation Memory utilization Memory saturation Network Interface utilization Storage Device I/O utilization Storage Device I/O saturation Storage Device I/O errors Metric
The USE Method: Easy Combinations Resource Type Metric CPU utilization CPU utilization CPU saturation run-queue length Memory utilization Memory saturation paging or swapping Network Interface utilization Storage Device I/O utilization available memory RX/TX tput/bandwidth device busy percent Storage Device I/O saturation wait queue length Storage Device I/O errors device errors
The USE Method: Harder Combinations Resource Type CPU errors Network saturation Storage Controller utilization CPU Interconnect utilization Mem. Interconnect saturation I/O Interconnect saturation Metric
The USE Method: Harder Combinations Resource Type Metric CPU errors eg, correctable CPU cache ECC events Network saturation "nocanputs", buffering Storage Controller utilization CPU Interconnect utilization active vs max controller IOPS and tput per port tput / max bandwidth Mem. Interconnect saturation memory stall cycles I/O Interconnect bus throughput / max saturation bandwidth
The USE Method: tools • To be thorough, you will need to use: • CPU performance counters • For bus and interconnect activity; eg, perf events, cpustat • Dynamic Tracing • For missing saturation and error metrics; eg, DTrace • Both can get tricky; tools can be developed to help • Please, no more top variants! ... unless it is interconnect-top or bus-top • I’ve written dozens of open source tools for both CPC and DTrace; much more can be done
Workload Characterization • May use as a 3rd Methodology • Characterize workload by: • who is causing the load? PID, UID, IP addr, ... • why is the load called? code path • what is the load? IOPS, tput, type • how is the load changing over time? • Best performance wins are from eliminating unnecessary work • Identifies class of issues that are load-based, not architecture-based
Drill-Down Analysis • May use as a 4th Methodology • Peel away software layers to drill down on the issue • Eg, software stack I/O latency analysis: Application System Call Interface File System Block Device Interface Storage Device Drivers Storage Devices
Drill-Down Analysis: Open Source • With Dynamic Tracing, all function entry & return points can be traced, with nanosecond timestamps. • One Strategy is to measure latency pairs, to search for the source; eg, A->B & C->D: static int arc_cksum_equal(arc_buf_t *buf) A{ zio_cksum_t zc; int equal; C mutex_enter(&buf->b_hdr->b_freeze_lock); fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); D equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); mutex_exit(&buf->b_hdr->b_freeze_lock); B} return (equal);
Other Methodologies • Method R • A latency-based analysis approach for Oracle databases. See "Optimizing Oracle Performance" by Cary Millsap and Jeff Holt (2003) • Experimental approaches • Can be very useful: eg, validating network throughput using iperf
Specific Tools for the USE Method
illumos-based • http://dtrace.org/blogs/brendan/2012/03/01/the-usemethod-solaris-performance-checklist/ Resource Type Metric CPU Utilization per-cpu: mpstat 1, "idl"; system-wide: vmstat 1, "id"; per-process:prstat -c 1 ("CPU" == recent), prstat mLc 1 ("USR" + "SYS"); per-kernel-thread: lockstat -Ii rate, DTrace profile stack() Saturation system-wide: uptime, load averages; vmstat 1, "r"; DTrace dispqlen.d (DTT) for a better "vmstat r"; per-process: prstat -mLc 1, "LAT" Errors fmadm faulty; cpustat (CPC) for whatever error counters are supported (eg, thermal throttling) Saturation system-wide: vmstat 1, "sr" (bad now), "w" (was very bad); vmstat -p 1, "api" (anon page ins == pain), "apo"; per-process: prstat -mLc 1, "DFL"; DTrace anonpgpid.d (DTT), vminfo:::anonpgin on execname CPU CPU Memory • ... etc for all combinations (would span a dozen slides)
Linux-based • http://dtrace.org/blogs/brendan/2012/03/07/the-usemethod-linux-performance-checklist/ Resource Type Metric CPU Utilization per-cpu: mpstat -P ALL 1, "%idle"; sar -P ALL, "%idle"; system-wide: vmstat 1, "id"; sar -u, "%idle"; dstat -c, "idl"; per-process:top, "%CPU"; htop, "CPU%"; ps -o pcpu; pidstat 1, "%CPU"; per-kernel-thread: top/htop ("K" to toggle), where VIRT == 0 (heuristic). [1] Saturation system-wide: vmstat 1, "r" > CPU count [2]; sar -q, "runq-sz" > CPU count; dstat -p, "run" > CPU count; perprocess: /proc/PID/schedstat 2nd field (sched_info.run_delay); perf sched latency (shows "Average" and "Maximum" delay per-schedule); dynamic tracing, eg, SystemTap schedtimes.stp "queued(us)" [3] Errors perf (LPE) if processor specific error events (CPC) are available; eg, AMD64′s "04Ah Single-bit ECC Errors Recorded by Scrubber" [4] CPU CPU • ... etc for all combinations (would span a dozen slides)
Products • Earlier I said methodologies could be supported by monitoring solutions • At Joyent we develop Cloud Analytics:
Future • Methodologies for advanced performance issues • I recently worked a complex KVM bandwidth issue where no current methodologies really worked • Innovative methods based on open source + dynamic tracing • Less performance mystery. Less guesswork. • Better use of resources (price/performance) • Easier for beginners to get started
Thank you • Resources: • http://dtrace.org/blogs/brendan • http://dtrace.org/blogs/brendan/2012/02/29/the-use-method/ • http://dtrace.org/blogs/brendan/tag/usemethod/ • http://dtrace.org/blogs/brendan/2011/12/18/visualizing-deviceutilization/ - ideas if you are a monitoring solution developer • brendan@joyent.com