Talk for AWS re:Invent 2014. Video: https://www.youtube.com/watch?v=7Cyd22kOqWc . Netflix tunes Amazon EC2 instances for maximum performance. In this session, you learn how Netflix configures the fastest possible EC2 instances, while reducing latency outliers. This session explores the various Xen modes (e.g., HVM, PV, etc.) and how they are optimized for different workloads. Hear how Netflix chooses Linux kernel versions based on desired performance characteristics and receive a firsthand look at how they set kernel tunables, including hugepages. You also hear about Netflix’s use of SR-IOV to enable enhanced networking and their approach to observability, which can exonerate EC2 issues and direct attention back to application performance.

1. PFC306
Brendan Gregg, Performance Engineering, Netflix
November 12, 2014 | Las Vegas, NV

9. S3
EC2
Cassandra
Applications
(Services)
EVCache
ELB
Elasticsearch
SES SQS

13. Start
i2 Select memory to
cache working set
Find best
balance

14. ASG-v011
…
Instance
Instance
Instance
ASG Cluster
prod1
ASG-v010
…
Instance
Instance
Instance
Canary
ELB

17. Select instance families Select resources
From any desired
resource, see
types & cost

18. eg, 8 vCPU:

21. Acceptable Headroom Unacceptable

26. Cost per hour
Services

36. # schedtool –B PID

37. vm.swappiness = 0 # from 60

38. # echo never > /sys/kernel/mm/transparent_hugepage/enabled # from madvise

39. vm.dirty_ratio = 80 # from 40
vm.dirty_background_ratio = 5 # from 10
vm.dirty_expire_centisecs = 12000 # from 3000
mount -o defaults,noatime,discard,nobarrier …

40. /sys/block/*/queue/rq_affinity2
/sys/block/*/queue/scheduler noop
/sys/block/*/queue/nr_requests256
/sys/block/*/queue/read_ahead_kb 256
mdadm –chunk=64 ...

41. net.core.somaxconn = 1000
net.core.netdev_max_backlog = 5000
net.core.rmem_max = 16777216
net.core.wmem_max = 16777216
net.ipv4.tcp_wmem = 4096 12582912 16777216
net.ipv4.tcp_rmem = 4096 12582912 16777216
net.ipv4.tcp_max_syn_backlog = 8096
net.ipv4.tcp_slow_start_after_idle = 0
net.ipv4.tcp_tw_reuse = 1
net.ipv4.ip_local_port_range = 10240 65535
net.ipv4.tcp_abort_on_overflow = 1 # maybe

42. echo tsc > /sys/devices/system/clocksource/clocksource0/current_clocksource

47. Resource
Utilization
X (%)

51. Application
System Libraries
System Calls
Kernel
Devices

54. $ sar -n TCP,ETCP,DEV 1
Linux 3.2.55 (test-e4f1a80b) 08/18/2014 _x86_64_ (8 CPU)
09:10:43 PM IFACE rxpck/s txpck/s rxkB/s txkB/s rxcmp/s txcmp/s rxmcst/s
09:10:44 PM lo 14.00 14.00 1.34 1.34 0.00 0.00 0.00
09:10:44 PM eth0 4114.00 4186.00 4537.46 28513.24 0.00 0.00 0.00
09:10:43 PM active/s passive/s iseg/s oseg/s
09:10:44 PM 21.00 4.00 4107.00 22511.00
09:10:43 PM atmptf/s estres/s retrans/s isegerr/s orsts/s
09:10:44 PM 0.00 0.00 36.00 0.00 1.00
[…]

59. Stack frame
Mouse-over
frames to
quantify
Ancestry

60. # git clone https://github.com/brendangregg/FlameGraph
# cd FlameGraph
# perf record -F 99 -ag -- sleep 60
# perf script | ./stackcollapse-perf.pl | ./flamegraph.pl > perf.svg

62. Broken
Java stacks
(missing
frame
pointer)
Kernel
TCP/IP
GC
Idle
thread
Time
Locks
epoll

65. # ./iosnoop –ts
Tracing block I/O. Ctrl-C to end.
STARTs ENDs COMM PID TYPE DEV BLOCK BYTES LATms
5982800.302061 5982800.302679 supervise 1809 W 202,1 17039600 4096 0.62
5982800.302423 5982800.302842 supervise 1809 W 202,1 17039608 4096 0.42
5982800.304962 5982800.305446 supervise 1801 W 202,1 17039616 4096 0.48
5982800.305250 5982800.305676 supervise 1801 W 202,1 17039624 4096 0.43
[…]
# ./iosnoop –h
USAGE: iosnoop [-hQst] [-d device] [-i iotype] [-p PID] [-n name] [duration]
-d device # device string (eg, "202,1)
-i iotype # match type (eg, '*R*' for all reads)
-n name # process name to match on I/O issue
-p PID # PID to match on I/O issue
-Q # include queueing time in LATms
-s # include start time of I/O (s)
-t # include completion time of I/O (s)
[…]

67. # perf record –e skb:consume_skb –ag -- sleep 10
# perf report
[...]
74.42% swapper [kernel.kallsyms] [k] consume_skb
|
--- consume_skb
arp_process
arp_rcv
__netif_receive_skb_core
__netif_receive_skb
netif_receive_skb
virtnet_poll
net_rx_action
__do_softirq
irq_exit
do_IRQ
ret_from_intr
[…]
Summarizing stack traces for a
tracepoint
perf_events can do many things,
it is hard to pick just one example

69. ec2-guest# ./showboost
CPU MHz : 2500
Turbo MHz : 2900 (10 active)
Turbo Ratio : 116% (10 active)
CPU 0 summary every 5 seconds...
Real CPU MHz
TIME C0_MCYC C0_ACYC UTIL RATIO MHz
06:11:35 6428553166 7457384521 51% 116% 2900
06:11:40 6349881107 7365764152 50% 115% 2899
06:11:45 6240610655 7239046277 49% 115% 2899
[...]

72. Region App Breakdowns
Metrics
Options
Interactive
Graph
Summary Statistics

75. Utilization Saturation
Errors
Per device
Breakdowns

78. http://aws.amazon.com/ec2/instance-types/
http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instance-types.html
http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/enhanced-networking.html
http://www.slideshare.net/cpwatson/cpn302-yourlinuxamioptimizationandperformance
http://www.brendangregg.com/blog/2014-09-27/from-clouds-to-roots.html
http://www.brendangregg.com/blog/2014-05-07/what-color-is-your-xen.html
http://www.brendangregg.com/linuxperf.html
http://www.slideshare.net/brendangregg/linux-performance-tools-2014
http://www.brendangregg.com/USEmethod/use-linux.html
http://www.brendangregg.com/blog/2014-06-12/java-flame-graphs.html
https://github.com/brendangregg/FlameGraph https://github.com/brendangregg/perf-tools

80. Talk Time Title
PFC-305 Wednesday, 1:15pm Embracing Failure: Fault Injection and Service Reliability
BDT-403 Wednesday, 2:15pm Next Generation Big Data Platform at Netflix
PFC-306 Wednesday, 3:30pm Performance Tuning EC2
DEV-309 Wednesday, 3:30pm From Asgard to Zuul, How Netflix’s proven Open Source
Tools can accelerate and scale your services
ARC-317 Wednesday, 4:30pm Maintaining a Resilient Front-Door at Massive Scale
PFC-304 Wednesday, 4:30pm Effective Inter-process Communications in the Cloud: The
Pros and Cons of Micro Services Architectures
ENT-209 Wednesday, 4:30pm Cloud Migration, Dev-Ops and Distributed Systems
APP-310 Friday, 9:00am Scheduling using Apache Mesos in the Cloud