OS caused Large JVM pauses: Deep dive and solutions

OS-caused Long JVM Pauses
- Deep Dive and Solutions
Zhenyun Zhuang
LinkedIn Corp., Mountain View, California, USA
https://www.linkedin.com/in/zhenyun
Zhenyun@gmail.com

Outline
 Introduction
 Background
 Scenario 1: startup state
 Scenario 2: steady state with memory pressure
 Scenario 3: steady state with heavy IO
 Lessons learned
2

Introduction
 Java + Linux
 Java is popular in production deployments
 Linux features interact with JVM operations
 Unique challenges caused by concurrent applications
 Long JVM pauses caused by Linux OS
 Production issues, in three scenarios
 Root causes
 Solutions
 References
 Ensuring High-performance of Mission-critical Java Applications in Multi-
tenant Cloud Platforms, IEEE Cloud 2014
 Eliminating Large JVM GC Pauses Caused by Background IO Traffic,
LinkedIn Engineering Blog, 2016 (Too many tweets bringing down a twitter server! :)
3

Background
 JVM and Heap
 Oracle HotSpot JVM
 Garbage collection
 Generations
 Garbage collectors
 Linux OS
 Paging (Regular page, Huge page)
 Swapping (Anonymous memory)
 Page cache writeback (Batched, Periodic)
4

Scenarios
 Three scenarios
 Startup state
 Steady state with memory pressure
 Steady state with heavy IO
 Workload
 Java application keeps allocating/de-allocating objects
 Background applications taking memories or issuing disk IO
 Performance metrics
 Application throughput (K allocations/sec)
 Java GC pauses
5

Scenario 1: Startup State (App. Symptoms)
 When Java applications start
 Life is good in the beginning
 Then Java throughput drops sharply
 Java GC pauses spike during the same period
6

Scenario 1: Startup State (Investigations)
 Java heap is gradually allocated
 Without enough memory, direct
page scanning can happen
 Heap is swapped out and in
 It causes large GC
7

Solutions
 Pre-allocating JVM heap spaces
 JVM “-XX:AlwaysPreTouch”
 Protecting JVM heap spaces from being
swapped out
 Swappoff command
 Swappiness
• =0 for kernel version before 2.6.32-303
• =1 for kernel version from 2.6.32-303
 Cgroup
8

Evaluations (Pre-allocating Heap)
9

Evaluations (Protecting Heap)
18 24
10

Scenario 2: Steady State (App. Symptoms)
 During steady state of a Java application, system
memory stresses due to other applications
 Java throughput drops sharply and performs badly
 Java GC pauses spike
11

Scenario 2: Steady State (Level-1 Investigations)
 During GC pauses, swapping activities persist
 Swapping in JVM pages causes GC pauses
 However, swapping is not enough
 Excessive GC pauses (i.e., 55 seconds)
 High sys-cpu usage (swapping is not sys-cpu intensive)
12
[Times: user=0.12 sys=54.67,
real=54.83 secs]

Scenario 2: Steady State (Level-2 Investigations)
 THP (Transparent Huge Pages)
 Improved TLB cache-hits
 Bi-directional operations
 THPs are allocated first, but split during memory pressure
 Regular pages are collapsed to make THPs
 CPU heavy, and thrashing!
4KB
Regular Pages
4KB
4KB
4KB
4KB
4KB
……
……
2MB
Transparent Huge
Pages (THP)
Splitting
Collapsing
13

Solutions
 Dynamically adjusting THP
 Enable THP when no memory pressure
 Disable THP during memory pressure period
 Fine tuning of THP parameters
14

Evaluations (Dynamic THP)
 Without memory pressure
 Dynamic THP delivers similar performance as THP is on
Mechanism THP Off THP On Dynamic THP
Throughput
(K allocations/sec)
12 15 15
Mechanism THP Off THP On Dynamic THP
Throughput
(K allocations/sec)
13 11 12
 With memory pressure
 Dynamic THP has some performance overhead
 Performance is less than THP-off
 But better than THP-on
15

Scenario 3: Steady State (Heavy IO)
 Production issue
 Online products
 Applications have light workload
 Both CMS and G1 garbage collectors
 Preliminary investigations
 Examined many layers/metrics
 The only suspect: disk IO occasionally is heavy
 But all application IO are asynchronous
16

Reproducing the problem
 Workload
 Simplified to avoid complex business logic
 https://github.com/zhenyun/JavaGCworkload
 Background IO
 Saturating HDD
17

Case I: Without background IO
18
No single longer-than-200ms pause

Case II: With background IO
Huge pause!
19

Time lines
 At time 35.04 (line 2), a young GC starts and takes 0.12 seconds to complete.
 The young GC finishes at time 35.16 and JVM tries to output the young GC statistics
to gc log file by issuing a write() system call (line 4).
 The write() call finishes at time 36.64 after being blocked for 1.47 seconds (line 5)
 When write() call returns to JVM, JVM records at time 36.64 this STW pause of 1.59
seconds (i.e., 0.12 + 1.47) (line 3).
21

Interaction between JVM and OS
22

Non-blocking IO can be blocked
 Stable page write
 For file-backed writing, OS writes to page cache first
 OS has write-back mechanism to persist dirty pages
 If a page is under write-back, the page is locked
 Journal committing
 Journals are generated for journaling file system
 When appending GC log files needs new blocks,
journals need to be committed
 Commitment might need to wait
23

Background IO activities
 OS activity such as swapping
 Data writing to underlying disks
 Administration and housekeeping software
 System-level software such as CFEngine also perform
disk IO
 Other co-located applications
 Co-located applications that share the disk drives,
then other applications contend on IO
 IO of the same JVM instance
 The particular JVM instance may use disk IO in ways
other than GC logging
24

Solutions
 Enhancing JVM
 Another thread
 Exposing JVM flags
 Reducing IO activities
 OS, other apps, same app
 Latency sensitive applications
 Separate disk
 High performing disks such as SSD
 Tmpfs
25

Evaluation
 SSD as the disk
26

The good, the bad, and the ugly
 The good: low real time
 Low user time and low sys time
 [user=0.18 sys=0.01, real=0.04 secs]
 The bad: non-low (but not high) real time
 High user time and low sys time
 [user=8.00 sys=0.02, real=0.50 secs]
 The ugly: high real time
 High sys time [user=0.02 sys=1.20, real=1.20 secs]
 Low sys time, low user time [Example? ]
27

Lessons Learned (I)
 Be cautious about Linux’s (and other OS) new
features
 Constantly incorporating new features to optimize
performance
 Some features incur performance tradeoff
 They may backfire in certain scenarios
28

Lessons Learned (II)
29
 Root causes can come from seemingly
insignificant information
 Linux emits significant amount of performance
information
 Most of us most of the time mostly only examine a
small subset of them
 Don’t ignore others – understand the interactions
of sub-components

Lessons Learned (III)
30
 Pay attention to multi-layer interaction
 Application protocol, JVM, OS, storage/networking
 Most people are familiar with a few layers
 Optimizations done at one layer may adversely
affect other layers
 Many performance problems are caused by the
cross-layer interactions

OS caused Large JVM pauses: Deep dive and solutions

More Related Content

What's hot (20)

Similar to OS caused Large JVM pauses: Deep dive and solutions (20)

More from Zhenyun Zhuang (20)

Recently uploaded (20)

OS caused Large JVM pauses: Deep dive and solutions