Nov 2015

Brendan Gregg
Senior Performance Architect, Netflix
CAUTION: PERFORMANCE TOOLS
• Over 60 million subscribers
• AWS EC2 Linux cloud
• FreeBSD CDN
• Awesome place to work
This Talk
• Observability, benchmarking, anti-patterns, and lessons
• Broken and misleading things that are surprising

Note: problems with current implementations are discussed, which

may be fixed/improved in the future
Observability:
System Metrics
LOAD AVERAGES
RFC 546
Load Averages (1, 5, 15 min)
$ uptime
22:08:07 up 9:05, 1 user, load average: 11.42, 11.87, 12.12

• "load"
– Usually CPU demand (scheduler run queue length/latency)
– On Linux, task demand: CPU + uninterruptible disk I/O (?)
• "average"
– Exponentially damped moving sum
• "1, 5, and 15 minutes"
– Constants used in the equation
• Don't study these for longer than 10 seconds
t=0 1
Load begins
(1 thread)

@ 1 min:
1 min avg =~ 0.62 5

15
Load Average

"1 minute load average"

really means…

"The exponentially damped moving sum of

CPU + uninterruptible disk I/O that uses
a value of 60 seconds in its equation"
TOP %CPU
top %CPU
$ top - 20:15:55 up 19:12, 1 user, load average: 7.96, 8.59, 7.05
Tasks: 470 total, 1 running, 468 sleeping, 0 stopped, 1 zombie
%Cpu(s): 28.1 us, 0.4 sy, 0.0 ni, 71.2 id, 0.0 wa, 0.0 hi, 0.1 si, 0.1 st
KiB Mem: 61663100 total, 61342588 used, 320512 free, 9544 buffers
KiB Swap: 0 total, 0 used, 0 free. 3324696 cached Mem

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND

11959 apiprod 20 0 81.731g 0.053t 14476 S 935.8 92.1 13568:22 java
12595 snmp 20 0 21240 3256 1392 S 3.6 0.0 2:37.23 snmp-pass
10447 snmp 20 0 51512 6028 1432 S 2.0 0.0 2:12.12 snmpd
18463 apiprod 20 0 23932 1972 1176 R 0.7 0.0 0:00.07 top
[…]

• Who is consuming CPU?

• And by how much?
top: Missing %CPU
• Short-lived processes can be missing entirely
– Process creates and exits in-between sampling /proc.
e.g., software builds.
– Try atop(1), or sampling using perf(1)
• Stop clearing the screen!
– No option to turn this off. Your eyes can miss updates.
– I often use pidstat(1) on Linux instead. Scroll back for history.
top: Misinterpreting %CPU
• Different top(1)s use different calculations
- On different OSes, check the man page, and run a test!
• %CPU can mean:
– A) Sum of per-CPU percents (0-Ncpu x 100%) consumed
during the last interval
– B) Percentage of total CPU capacity (0-100%) consumed
during the last interval
– C) (A) but historically damped (like load averages)
– D) (B) " " "
top: %Cpu vs %CPU
$ top - 15:52:58 up 10 days, 21:58, 2 users, load average: 0.27, 0.53, 0.41
Tasks: 180 total, 1 running, 179 sleeping, 0 stopped, 0 zombie
%Cpu(s): 1.2 us, 24.5 sy, 0.0 ni, 67.2 id, 0.2 wa, 0.0 hi, 6.6 si, 0.4 st
KiB Mem: 2872448 total, 2778160 used, 94288 free, 31424 buffers
KiB Swap: 4151292 total, 76 used, 4151216 free. 2411728 cached Mem

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND

12678 root 20 0 96812 1100 912 S 100.4 0.0 0:23.52 iperf
12675 root 20 0 170544 1096 904 S 88.8 0.0 0:20.83 iperf
215 root 20 0 0 0 0 S 0.3 0.0 0:27.73 jbd2/sda1-8
[…]

• This 4 CPU system is consuming:

– 130% total CPU, via %Cpu(s)
– 190% total CPU, via %CPU
• Which one is right? Is either?
CPU Summary Statistics
• %Cpu row is from /proc/stat
• linux/Documentation/cpu-load.txt:

In most cases the `/proc/stat' information reflects

the reality quite closely, however due to the nature
of how/when the kernel collects this data
sometimes it can not be trusted at all.

• /proc/stat is used by everything for CPU stats

%CPU
What is %CPU anyway?
• "Good" %CPU:
– Retiring instructions (provided they aren't a spin loop)
– High IPC (Instructions-Per-Cycle)
• "Bad" %CPU:
– Stall cycles waiting on resources, usually memory I/O
– Low IPC
– Buying faster processors may make little difference
• %CPU alone is ambiguous
– Would love top(1) to split %CPU into cycles retiring vs stalled
– Although, it gets worse…
A CPU Mystery…
• As load increased, CPU ms per request lowered (blue)
– up to 1.84x faster
• Was it due to:
- Cache warmth? no
- Different code? no
- Turbo boost? no
• (Same test, but
problem fixed, is
shown in red)
CPU Speed Variation
• Clock speed can vary thanks to:
– Intel Turbo Boost: by hardware, based on power, temp, etc
– Intel Speed Step: by software, controlled by the kernel
• %CPU is still ambiguous, given IPC. Need to know the
clock speed as well
• CPU counters nowadays have "reference cycles"
Out-of-order Execution
• CPUs execute uops out-of-
order and in parallel across
multiple functional units
• %CPU doesn't account for
how many units are active
• Accounting each cycles as
"stalled" or “retiring" is a
simplification
• Nowadays it's a lot of work
to truly understand what
CPUs are doing

https://upload.wikimedia.org/wikipedia/commons/6/64/Intel_Nehalem_arch.svg
I/O WAIT
I/O Wait
$ mpstat -P ALL 1
08:06:43 PM CPU %usr %nice %sys %iowait %irq %soft %steal %guest %idle
08:06:44 PM all 53.45 0.00 3.77 0.00 0.00 0.39 0.13 0.00 42.26
[…]

• Suggests system is disk I/O bound, but often misleading

• Comparing I/O wait between system A and B:
- higher might be bad: slower disks, more blocking
- lower might be bad: slower processor and architecture
consumes more CPU, obscuring I/O wait
• Solaris implementation was also broken and later
hardwired to zero
• Can be very useful when understood: another idle state
I/O Wait Venn Diagram
Per CPU:

CPU Waiting for disk I/O

"CPU" "CPU" "I/O Wait"

"Idle"
FREE MEMORY
Free Memory
$ free -m
total used free shared buffers cached
Mem: 3750 1111 2639 0 147 527
-/+ buffers/cache: 436 3313
Swap: 0 0 0

• "free" is near-zero: I'm running

out of memory!
- No, it's in the file system cache,
and is still free for apps to use
• Linux free(1) explains it, but
other tools, e.g. vmstat(1), don't
• Some file systems (e.g., ZFS)
may not be shown in the www.linuxatemyram.com
system's cached metrics at all
VMSTAT
vmstat(1)
$ vmstat –Sm 1
procs -----------memory---------- ---swap-- -----io---- -system-- ----cpu----
r b swpd free buff cache si so bi bo in cs us sy id wa
8 0 0 1620 149 552 0 0 1 179 77 12 25 34 0 0
7 0 0 1598 149 552 0 0 0 0 205 186 46 13 0 0
8 0 0 1617 149 552 0 0 0 8 210 435 39 21 0 0
8 0 0 1589 149 552 0 0 0 0 218 219 42 17 0 0
[…]

• Linux: first line has some summary since boot values —

confusing!
• Other implementations:
- "r" may be sampled once per second. Almost useless.
- Columns like "de" for deficit, making much less sense for non-
page scanned situations
NETSTAT -S
netstat -s
$ netstat -s 1428960 acknowledgments not containing data received
Ip: 1004791 predicted acknowledgments
7962754 total packets received 1 times recovered from packet loss due to fast retransmit
8 with invalid addresses 5044 times recovered from packet loss due to SACK data
0 forwarded 2 bad SACKs received
0 incoming packets discarded Detected reordering 4 times using SACK
7962746 incoming packets delivered Detected reordering 11 times using time stamp
8019427 requests sent out 13 congestion windows fully recovered
Icmp: 11 congestion windows partially recovered using Hoe heuristic
382 ICMP messages received TCPDSACKUndo: 39
0 input ICMP message failed. 2384 congestion windows recovered after partial ack
ICMP input histogram: 228 timeouts after SACK recovery
destination unreachable: 125 100 timeouts in loss state
timeout in transit: 257 5018 fast retransmits
3410 ICMP messages sent 39 forward retransmits
0 ICMP messages failed 783 retransmits in slow start
ICMP output histogram: 32455 other TCP timeouts
destination unreachable: 3410 TCPLossProbes: 30233
IcmpMsg: TCPLossProbeRecovery: 19070
InType3: 125 992 sack retransmits failed
InType11: 257 18 times receiver scheduled too late for direct processing
OutType3: 3410 705 packets collapsed in receive queue due to low socket buffer
Tcp: 13658 DSACKs sent for old packets
17337 active connections openings 8 DSACKs sent for out of order packets
395515 passive connection openings 13595 DSACKs received
8953 failed connection attempts 33 DSACKs for out of order packets received
240214 connection resets received 32 connections reset due to unexpected data
3 connections established 108 connections reset due to early user close
7198375 segments received 1608 connections aborted due to timeout
7504939 segments send out TCPSACKDiscard: 4
62696 segments retransmited TCPDSACKIgnoredOld: 1
10 bad segments received. TCPDSACKIgnoredNoUndo: 8649
1072 resets sent TCPSpuriousRTOs: 445
InCsumErrors: 5 TCPSackShiftFallback: 8588
Udp: TCPRcvCoalesce: 95854
759925 packets received TCPOFOQueue: 24741
3412 packets to unknown port received. TCPOFOMerge: 8
0 packet receive errors TCPChallengeACK: 1441
784370 packets sent TCPSYNChallenge: 5
UdpLite: TCPSpuriousRtxHostQueues: 1
TcpExt: TCPAutoCorking: 4823
858 invalid SYN cookies received IpExt:
8951 resets received for embryonic SYN_RECV sockets InOctets: 1561561375
14 packets pruned from receive queue because of socket buffer overrun OutOctets: 1509416943
6177 TCP sockets finished time wait in fast timer InNoECTPkts: 8201572
293 packets rejects in established connections because of timestamp InECT1Pkts: 2
733028 delayed acks sent InECT0Pkts: 3844
89 delayed acks further delayed because of locked socket InCEPkts: 306
Quick ack mode was activated 13214 times
336520 packets directly queued to recvmsg prequeue.
43964 packets directly received from backlog
11406012 packets directly received from prequeue
1039165 packets header predicted
7066 packets header predicted and directly queued to user
netstat -s

[…]
Tcp:
17337 active connections openings
395515 passive connection openings
8953 failed connection attempts
240214 connection resets received
3 connections established
7198870 segments received
7505329 segments send out
62697 segments retransmited
10 bad segments received.
1072 resets sent
InCsumErrors: 5
[…]
netstat -s
• Many metrics on Linux (can be over 200)
• Still doesn't include everything: getting better, but don't
assume everything is there
• Includes typos & inconsistencies
• Might be more readable to:
cat /proc/net/snmp /proc/net/netstat
• Totals since boot can be misleading
• On Linux, -s needs -c support
• Often no documentation outside kernel source code
• Requires expertise to comprehend
DISK METRICS
Disk Metrics
• All disk metrics are misleading
• Disk %utilization / %busy
– Logical devices (volume managers) can process requests in
parallel, and may accept more I/O at 100%
• Disk IOPS
– High IOPS is "bad"? That depends…
• Disk latency
– Does it matter? File systems and volume managers try hard
to hide latency and make latency asynchronous
– Better measuring latency via application->FS calls
Rules for Metrics Makers
• They must work
– As well as possible. Clearly document caveats.
• They must be useful
– Document a real use case (eg, my example.txt files).
If you get stuck, it's not useful – ditch it.
• Aim to be intuitive
– Document it. If it's too weird to explain, redo it.
• As few as possible
– Respect end-user's time
• Good system examples:
– iostat -x: workload columns, then resulting perf columns
– Linux sar: consistency, units on columns, logical groups
Observability:
Profilers
PROFILERS
Linux perf
• Can sample stack traces and summarize output:
# perf report -n -stdio
[…]
# Overhead Samples Command Shared Object Symbol
# ........ ............ ....... ................. .............................
#
20.42% 605 bash [kernel.kallsyms] [k] xen_hypercall_xen_version
|
--- xen_hypercall_xen_version
check_events
|
|--44.13%-- syscall_trace_enter
| tracesys
| |
| |--35.58%-- __GI___libc_fcntl
| | |
| | |--65.26%-- do_redirection_internal
| | | do_redirections
| | | execute_builtin_or_function
| | | execute_simple_command
[… ~13,000 lines truncated …]
Too Much Output
… as a Flame Graph
PROFILER VISIBILITY
System Profilers with Java

Java Kernel
TCP/IP

JVM GC
Locks epoll
Idle
Time
thread
System Profilers with Java
• e.g., Linux perf
• Visibility
– JVM (C++)
– GC (C++)
– libraries (C)
– kernel (C)
• Typical problems (x86):
– Stacks missing for Java and other runtimes
– Symbols missing for Java methods
• Profile everything except Java and similar runtimes
Java Profilers

Kernel,
libraries,
JVM
Java

GC
Java Profilers
• Visibility
– Java method execution
– Object usage
– GC logs
– Custom Java context
• Typical problems:
– Sampling often happens at safety/yield points (skew)
– Method tracing has massive observer effect
– Misidentifies RUNNING as on-CPU (e.g., epoll)
– Doesn't include or profile GC or JVM CPU time
– Tree views not quick (proportional) to comprehend
• Inaccurate (skewed) and incomplete profiles
COMPILER OPTIMIZATIONS
Broken System Stack Traces
# perf record –F 99 –a –g – sleep 30
• Profiling Java on # perf script
[…]
x86 using perf java 4579 cpu-clock:
ffffffff8172adff tracesys ([kernel.kallsyms])
• The stacks are 7f4183bad7ce pthread_cond_timedwait@@GLIBC_2…

1 or 2 levels java 4579 cpu-clock:

7f417908c10b [unknown] (/tmp/perf-4458.map)
deep, and have
junk values java 4579 cpu-clock:
7f4179101c97 [unknown] (/tmp/perf-4458.map)

java 4579 cpu-clock:

7f41792fc65f [unknown] (/tmp/perf-4458.map)
a2d53351ff7da603 [unknown] ([unknown])

java 4579 cpu-clock:

7f4179349aec [unknown] (/tmp/perf-4458.map)

java 4579 cpu-clock:

7f4179101d0f [unknown] (/tmp/perf-4458.map)
[…]
Why Stacks are Broken
• On x86 (x86_64), hotspot uses the frame pointer register
(RBP) as general purpose
• This "compiler optimization" breaks stack walking
• Once upon a time, x86 had fewer registers, and this
made much more sense
• gcc provides -fno-omit-frame-pointer to avoid
doing this
– JDK8u60+ now has this as -XX:+PreserveFramePoiner
Missing Symbols
• Missing symbols may show up as hex; e.g., Linux perf:
71.79% 334 sed sed [.] 0x000000000001afc1
|
|--11.65%-- 0x40a447
| 0x40659a
| 0x408dd8
| 0x408ed1 broken
| 0x402689
| 0x7fa1cd08aec5

12.06% 62 sed sed [.] re_search_internal

|
--- re_search_internal
|
|--96.78%-- re_search_stub
| rpl_re_search
| match_regex
| do_subst not broken
| execute_program
| process_files
| main
| __libc_start_main
Fixing Symbols
• For applications, install debug symbol package
• For JIT'd code, Linux perf already looks for an
externally provided symbol file: /tmp/perf-PID.map
# perf script
Failed to open /tmp/perf-8131.map, continuing without symbols
[…]
java 8131 cpu-clock:
7fff76f2dce1 [unknown] ([vdso])
7fd3173f7a93 os::javaTimeMillis() (/usr/lib/jvm…
7fd301861e46 [unknown] (/tmp/perf-8131.map)
[…]

• Find for a way to create this for your runtime

INSTRUCTION PROFILING
Instruction Profiling
# perf annotate -i perf.data.noplooper --stdio
Percent | Source code & Disassembly of noplooper
--------------------------------------------------------
: Disassembly of section .text:
:
: 00000000004004ed <main>:
0.00 : 4004ed: push %rbp
0.00 : 4004ee: mov %rsp,%rbp
20.86 : 4004f1: nop
0.00 : 4004f2: nop
0.00 : 4004f3: nop
0.00 : 4004f4: nop
19.84 : 4004f5: nop
0.00 : 4004f6: nop
0.00 : 4004f7: nop
0.00 :
18.73 :
4004f8:
4004f9:
nop
nop
• Often broken nowadays due
0.00 : 4004fa: nop to skid, out-of-order
0.00 : 4004fb: nop
0.00 : 4004fc: nop execution, and sampling the
19.08 : 4004fd: nop resumption instruction
0.00 : 4004fe: nop
0.00 : 4004ff: nop • Better with PEBS support
0.00 : 400500: nop
21.49 : 400501: jmp 4004f1 <main+0x4>
Observability:
Overhead
TCPDUMP
tcpdump
$ tcpdump -i eth0 -w /tmp/out.tcpdump
tcpdump: listening on eth0, link-type EN10MB (Ethernet), capture size 65535 bytes
^C7985 packets captured
8996 packets received by filter
1010 packets dropped by kernel

• Packet tracing doesn't scale. Overheads:

– CPU cost of per-packet tracing (improved by [e]BPF)
• Consider CPU budget per-packet at 10/40/100 GbE
– Transfer to user-level (improved by ring buffers)
– File system storage (more CPU, and disk I/O)
– Possible additional network transfer
• Can also drop packets when overloaded
• You should only trace send/receive as a last resort
– I solve problems by tracing lower frequency TCP events
STRACE
strace
• Before:
$ dd if=/dev/zero of=/dev/null bs=1 count=500k
[…]
512000 bytes (512 kB) copied, 0.103851 s, 4.9 MB/s

• After:
$ strace -eaccept dd if=/dev/zero of=/dev/null bs=1 count=500k
[…]
512000 bytes (512 kB) copied, 45.9599 s, 11.1 kB/s

• 442x slower. This is worst case.

• strace(1) pauses the process twice for each syscall.
This is like putting metering lights on your app.
– "BUGS: A traced process runs slowly." – strace(1) man page
– Use buffered tracing / in-kernel counters instead, e.g. DTrace
DTRACE
DTrace
• Overhead often negligible, but not always
• Before:
# time wc systemlog
262600 2995200 23925200 systemlog
real 0m1.098s
user 0m1.085s
sys 0m0.012s

• After:
# time dtrace -n 'pid$target:::entry { @[probefunc] = count(); }' -c 'wc systemlog'
dtrace: description 'pid$target:::entry ' matched 3756 probes
262600 2995200 23925200 systemlog
[…]
real 7m2.896s
user 7m2.650s
sys 0m0.572s

• 384x slower. Fairly worst case: frequent pid probes.

Tracing Dangers
• Overhead potential exists for all tracers
– Overhead = event instrumentation cost X frequency of event
• Costs
– Lower: counters, in-kernel aggregations
– Higher: event dumps, stack traces, string copies, copyin/outs
• Frequencies
– Lower: process creation & destruction, disk I/O (usually), …
– Higher: instructions, functions in I/O hot path, malloc/free,
Java methods, …
• Advice
– < 10,000 events/sec, probably ok
– > 100,000 events/sec, overhead may start to be measurable
DTraceToolkit
• My own tools that can cause massive overhead:
– dapptrace/dappprof: can trace all native functions
– Java/j_flow.d, ...: can trace all Java methods with
+ExtendedDTraceProbes
# j_flow.d
C PID TIME(us) -- CLASS.METHOD
0 311403 4789112583163 -> java/lang/Object.<clinit>
0 311403 4789112583207 -> java/lang/Object.registerNatives
0 311403 4789112583323 <- java/lang/Object.registerNatives
0 311403 4789112583333 <- java/lang/Object.<clinit>
0 311403 4789112583343 -> java/lang/String.<clinit>
0 311403 4789112583732 -> java/lang/String$CaseInsensitiveComparator.<init>
0 311403 4789112583743 -> java/lang/String$CaseInsensitiveComparator.<init>
0 311403 4789112583752 -> java/lang/Object.<init>
[...]

• Useful for debugging, but should warn about overheads

VALGRIND
Valgrind
• A suite of tools including an extensive leak detector

"Your program will run much slower

(eg. 20 to 30 times) than normal"

– http://valgrind.org/docs/manual/quick-start.html

• To its credit it does warn the end user

JAVA PROFILERS
Java Profilers
• Some Java profilers have two modes:
– Sampling stacks: eg, at 100 Hertz
– Tracing methods: instrumenting and timing every method
• Method timing has been described as "highly accurate",
despite slowing the target by up to 1000x!
• Issues & advice already covered at QCon:
– Nitsan Wakart "Profilers are Lying Hobbitses" earlier today
– Java track tomorrow
Observability:
Monitoring
MONITORING
Monitoring
• By now you should recognize these pathologies:
– Let's just graph the system metrics!
• That's not the problem that needs solving
– Let's just trace everything and post process!
• Now you have one million problems per second
• Monitoring adds additional problems:
– Let's have a cloud-wide dashboard update per-second!
• From every instance? Packet overheads?
– Now we have billions of metrics!
Observability:
Statistics
STATISTICS
Statistics

"Then there is the man who drowned crossing a

stream with an average depth of six inches."

– W.I.E. Gates
Statistics
• Averages can be misleading
– Hide latency outliers
– Per-minute averages can hide multi-second issues
• Percentiles can be misleading
– Probability of hitting 99.9th latency may be more than 1/1000
after many dependency requests
• Show the distribution:
– Summarize: histogram, density plot, frequency trail
– Over-time: scatter plot, heat map
• See Gil Tene's "How Not to Measure Latency" QCon talk
from earlier today
Average Latency
• When the index of central tendency isn't…
Observability:
Visualizations
VISUALIZATIONS
Tachometers

…especially with arbitrary color highlighting

Pie Charts

usr sys wait idle

…for real-time metrics

Doughnuts

usr sys wait idle

…like pie charts but worse

Traffic Lights

RED == BAD (usually)

GREEN == GOOD (hopefully)

…when used for subjective metrics

These can be used for objective metrics
Benchmarking
BENCHMARKING
~100% of benchmarks are wrong
"Most popular benchmarks are flawed"
Source: Traeger, A., E. Zadok, N. Joukov, and C. Wright.
“A Nine Year Study of File System and Storage
Benchmarking,” ACM Transactions on Storage, 2008.

Not only can a popular benchmark be broken, but so can

all alternates.
REFUTING BENCHMARKS
The energy needed
to refute benchmarks
is multiple orders of magnitude
bigger than to run them

It can take 1-2 weeks of senior performance engineering

time to debug a single benchmark.
Benchmarking
• Benchmarking is a useful form of experimental analysis
– Try observational first; benchmarks can perturb
• Accurate and realistic benchmarking is vital for technical
investments that improve our industry
• However, benchmarking is error prone
COMMON MISTAKES
Common Mistakes
1. Testing the wrong target
– eg, FS cache instead of disk; misconfiguration
2. Choosing the wrong target
– eg, disk instead of FS cache … doesn’t resemble real world
3. Invalid results
– benchmark software bugs
4. Ignoring errors
– error path may be fast!
5. Ignoring variance or perturbations
– real workload isn't steady/consistent, which matters
6. Misleading results
– you benchmark A, but actually measure B, and conclude you
measured C
PRODUCT EVALUATIONS
Product Evaluations
• Benchmarking is used for product evaluations & sales
• The Benchmark Paradox:
– If your product’s chances of winning a benchmark are
50/50, you’ll usually lose
– To justify a product switch, a customer may run several
benchmarks, and expect you to win them all
– May mean winning a coin toss at least 3 times in a row
– http://www.brendangregg.com/blog/2014-05-03/the-benchmark-paradox.html

• Solving this seeming paradox (and benchmarking):

– Confirm benchmark is relevant to intended workload
– Ask: why isn't it 10x?
Active Benchmarking
• Root cause performance analysis while the
benchmark is still running
– Use observability tools
– Identify the limiter (or suspected limiter) and include it with the
benchmark results
– Answer: why not 10x?
• This takes time, but uncovers most mistakes
MICRO BENCHMARKS
Micro Benchmarks
• Test a specific function in isolation. e.g.:
– File system maximum cached read operations/sec
– Network maximum throughput
• Examples of bad microbenchmarks:
– gitpid() in a tight loop
– speed of /dev/zero and /dev/null
• Common problems:
– Testing a workload that is not very relevant
– Missing other workloads that are relevant
MACRO BENCHMARKS
Macro Benchmarks
• Simulate application user load. e.g.:
– Simulated web client transaction
• Common problems:
– Misplaced trust: believed to be realistic, but misses variance,
errors, perturbations, e.t.c.
– Complex to debug, verify, and root cause
KITCHEN SINK BENCHMARKS
Kitchen Sink Benchmarks
• Run everything!
– Mostly random benchmarks found on the Internet, where
most are are broken or irrelevant
– Developers focus on collecting more benchmarks than
verifying or fixing the existing ones
• Myth that more benchmarks == greater accuracy
– No, use active benchmarking (analysis)
AUTOMATION
Automated Benchmarks
• Completely automated procedure. e.g.:
– Cloud benchmarks: spin up an instance, benchmark, destroy.
Automate.
• Little or no provision for debugging
• Automation is only part of the solution
Benchmarking:
More Examples
BONNIE++
bonnie++
• "simple tests of hard drive and file system performance"
• First metric printed by (thankfully) older versions:
per character sequential output
• What was actually tested:
– 1 byte writes to libc (via putc())
– 4 Kbyte writes from libc -> FS (depends on OS; see setbuffer())
– 128 Kbyte async writes to disk (depends on storage stack)
– Any file system throttles that may be present (eg, ZFS)
– C++ code, to some extent (bonnie++ 10% slower than Bonnie)
• Actual limiter:
– Single threaded write_block_putc() and putc() calls
APACHE BENCH
Apache Bench
• HTTP web server benchmark
• Single thread limited (use wrk for multi-threaded)
• Keep-alive option (-k):
– without: Can become an unrealistic TCP session benchmark
– with: Can become an unrealistic server throughput test
• Performance issues of ab's own code
UNIXBENCH
UnixBench
• The original kitchen-sink micro benchmark from 1984,
published in BYTE magazine
• Innovative & useful for the time, but that time has passed
• More problems than I can shake a stick at
• Starting with…
COMPILERS
UnixBench Makefile
• Default (by ./Run) for Linux. Would you edit it? Then what?
## Very generic
#OPTON = -O

## For Linux 486/Pentium, GCC 2.7.x and 2.8.x

#OPTON = -O2 -fomit-frame-pointer -fforce-addr -fforce-mem -ffast-math \
# -m486 -malign-loops=2 -malign-jumps=2 -malign-functions=2

## For Linux, GCC previous to 2.7.0

#OPTON = -O2 -fomit-frame-pointer -fforce-addr -fforce-mem -ffast-math -m486

#OPTON = -O2 -fomit-frame-pointer -fforce-addr -fforce-mem -ffast-math \

# -m386 -malign-loops=1 -malign-jumps=1 -malign-functions=1

## For Solaris 2, or general-purpose GCC 2.7.x

OPTON = -O2 -fomit-frame-pointer -fforce-addr -ffast-math -Wall

## For Digital Unix v4.x, with DEC cc v5.x

#OPTON = -O4
#CFLAGS = -DTIME -std1 -verbose -w0
UnixBench Makefile
• "Fixing" the Makefile improved the first result, Dhrystone
2, by 64%
• Is everyone "fixing" it the same way, or not? Are they
using the same compiler version? Same OS? (No.)
UnixBench Documentation

"The results will depend not only on your

hardware, but on your operating system,
libraries, and even compiler."

"So you may want to make sure that all your

test systems are running the same version of
the OS; or at least publish the OS and
compuiler versions with your results."
SYSTEM MICROBENCHMARKS
UnixBench Tests
• Results summarized as "The BYTE Index". From USAGE:
system:
dhry2reg Dhrystone 2 using register variables
whetstone-double Double-Precision Whetstone
syscall System Call Overhead
pipe Pipe Throughput
context1 Pipe-based Context Switching
spawn Process Creation
execl Execl Throughput
fstime-w File Write 1024 bufsize 2000 maxblocks
fstime-r File Read 1024 bufsize 2000 maxblocks
fstime File Copy 1024 bufsize 2000 maxblocks
fsbuffer-w File Write 256 bufsize 500 maxblocks
fsbuffer-r File Read 256 bufsize 500 maxblocks
fsbuffer File Copy 256 bufsize 500 maxblocks
fsdisk-w File Write 4096 bufsize 8000 maxblocks
fsdisk-r File Read 4096 bufsize 8000 maxblocks
fsdisk File Copy 4096 bufsize 8000 maxblocks
shell1 Shell Scripts (1 concurrent) (runs "looper 60 multi.sh 1")
shell8 Shell Scripts (8 concurrent) (runs "looper 60 multi.sh 8")
shell16 Shell Scripts (8 concurrent) (runs "looper 60 multi.sh 16")

• What can go wrong? Everything.

Anti-Patterns
ANTI-PATTERNS
Street Light Anti-Method
1. Pick observability tools that are:
– Familiar
– Found on the Internet
– Found at random
2. Run tools
3. Look for obvious issues
Blame Someone Else Anti-Method
1. Find a system or environment component you are not
responsible for
2. Hypothesize that the issue is with that component
3. Redirect the issue to the responsible team
4. When proven wrong, go to 1
Performance Tools Team
• Having a separate performance tools team, who creates
tools but doesn't use them (no production exposure)
• At Netflix:
– The performance engineering team builds tools and uses
tools for both service consulting and live production triage
• Mogul, Vector, …
– Other teams (CORE, traffic, …) also build performance tools
and use them during issues
• Good performance tools are built out of necessity
Messy House Fallacy
• Fallacy: my code is a mess, I bet yours is immaculate,
therefore the bug must be mine

• Reality: everyone's code is terrible and buggy

• When analyzing performance, don't overlook the system:

kernel, libraries, etc.
Lessons
PERFORMANCE TOOLS
Observability
• Trust nothing, verify everything
– Cross-check with other observability tools
– Write small "known" workloads, and confirm metrics match
– Find other sanity tests: e.g. check known system limits
– Determine how metrics are calculated, averaged, updated
• Find metrics to solve problems
– Instead of understanding hundreds of system metrics
– What problems do you want to observe? What metrics would
be sufficient? Find, verify, and use those. e.g., USE Method.
– The metric you want may not yet exist
• File bugs, get these fixed/improved
Observe Everything
• Use functional diagrams to pose Q's and find missing metrics:
Profile Everything
Java Mixed-Mode Flame Graph

Kernel
Java JVM
GC
Visualize Everything
Benchmark Nothing
• Trust nothing, verify everything
• Do Active Benchmarking:
1. Configure the benchmark to run in steady state, 24x7
2. Do root-cause analysis of benchmark performance
3. Answer: why is it not 10x?
Links & References
• https://www.rfc-editor.org/rfc/rfc546.pdf
• https://upload.wikimedia.org/wikipedia/commons/6/64/Intel_Nehalem_arch.svg
• http://www.linuxatemyram.com/
• Traeger, A., E. Zadok, N. Joukov, and C. Wright. “A Nine Year Study of File System
and Storage Benchmarking,” ACM Trans- actions on Storage, 2008.
• http://www.brendangregg.com/blog/2014-06-09/java-cpu-sampling-using-hprof.html
• http://www.brendangregg.com/blog/2014-05-03/the-benchmark-paradox.html
• http://www.brendangregg.com/ActiveBenchmarking/bonnie++.html
• https://blogs.oracle.com/roch/entry/decoding_bonnie
• http://www.brendangregg.com/blog/2014-05-02/compilers-love-messing-with-
benchmarks.html
• https://code.google.com/p/byte-unixbench/
• https://qconsf.com/sf2015/presentation/how-not-measure-latency
• https://qconsf.com/sf2015/presentation/profilers-lying
• Caution signs drawn be me, inspired by real-world signs
THANKS
Nov 2015

Thanks
• Questions?
• http://techblog.netflix.com
• http://slideshare.net/brendangregg
• http://www.brendangregg.com
• [email protected]
• @brendangregg