Java & low latency applications

Code & Coffee #22
Low-latency Java:
breaking the boundaries
Ruslan Shevchenko
<ruslan@shevchenko.kiev.ia>
@rssh1

Overview
❖ What is low-latency and when it needed.
❖ LL-speciﬁc programming techniques:
❖ Concurrent ﬂows
❖ Memory
❖ Layout issues,
❖ GC
❖ Unmanaged memory access.
❖ JNI

HPC != HPC
throughput
efﬁciency
speed
compromise
//low latency

low-latency java:
❖ Ineffective
❖ hardware is underused.
❖ Non-scalable
❖ operational range is strictly deﬁned.
❖ Uncomfortable
❖ api reﬂect low-level details.
// needed: soft realtime

soft realtime
❖ trading (HTF)
❖ video/audio processing:
❖ orchestration infrastructure:
// NYSE trading ﬂoor.
// was closed for humans in 200

softrealtime: why JVM (?)
❖ Balance (99 % - ‘Usual code’, 1% -
soft realtime)
❖ crossing JVM boundaries is expensive.
❖ don’t do this. [if you can]

Softrealtime & JVM: Programming techniques.
❖ Don’t guess, know.
❖ Measure
❖ profiling [not alw. practical];
❖ sampling (jvm option: -hprof) [jvisualvm, jconsole, flight recorder]
❖ log safepoint / gc pauses
❖ Experiments
❖ benchmarks: http://openjdk.java.net/projects/code-tools/jmh/
❖ critical patch: N2N flow.
❖ Know details, how code is executed.

Programming techniques: Concurrent models
❖ Concurrent flows:
❖ minimise context switches:
❖ don’t switch [ number of flows < number of all
processor cores].
❖ Java: main-thread + service-thread + listener-
thread; [ user threads: cores-3 ]
❖ pin thread to core [thread affinity]
❖ JNI [JNA], exists OSS libraries [OpenHFT]

Programming techniques: data exchange
❖ Concurrent ﬂows:
❖ data exchange between threads:
❖ each thread have own ‘local’ memory state.
// state can be unsynchronised
// local access is match faster than volatile
L1 cache: 0.3ns
L2 cache: 1ns
L3 cache: 10ns
RAM: 120ns
Volatile write = RAM write + invalidate cache lines.
❖ queue between threads is faster then shared access.
❖ locks are evil.

Programming techniques: memory issues
❖ GC
❖ young generation (relative short)
❖ full.
❖ [Stop-the-world], pause ~ heap size
❖ Contention
❖ ﬂushing cache lines is expensive

Soft-realtime: memory
❖ Object: what is java object for the JVM ?
class Order
{
long id
int timestamp;
String symbol;
long price;
int quantity;
Side side;
boolean limit;
Party party;
Venue venue;
}
1002,t,”APPL”,p,q,Bu
y/Sell, Owner,
Venue.

❖ Object: what is java object for the JVM ?
class Order
{
long id;
int timestamp;
String symbol;
long price;
int quantity;
Side side;
boolean limit;
Party party;
Venue venue;
}
t,”APPL”,p,q,Buy/
Sell, Owner, Venue.
header [64]:
class [32|64]
id [64]:
timestamp 32:
symbol [32-64];
price [64]
quantity[32]
side [32|64]
limit [32]
party [32|64]
venue [32|64]
… padding [0-63]
header,class:128
data: 64 “APPL”
header,class: 128
………..
header,class:128
………..
pointer (direct or compressed)
direct = address in memory

❖ Object: what JMV do ?
Order order = new Order() // alloc 80 bytes from heap
// order in a local variables (GC root)
int t = order.getTimestamp(); // memory read
Party p = order.getParty(); // memory read
order.setParty(p1);
// memory write +
//mark object to be rechecked in next GC cycle
//can force GC is memory reached threshold.

❖ How ‘idiomatic’ Java code will behave ?
List<Order> send = new ArrayList<>();
while(!buffer.isEmpty()) {
Order order = buffer.readOrder();
int fullPrice = order.price*order.quantity;
if (order.getParty().getLimit()+fullPrice > maxLimit) {
client.rejected(order,”party-limit”)
}else{
engine.process(order);
send.add(order);
}
}

❖ How ‘idiomatic’ Java code will behave ?
List<Order> send = new ArrayList<>();
}else{
send.add(order);
}
}
- order will be created
- added to long-lived collections
will cause ‘StopTheWorld’

❖ V1: Die young or live forever.
LongSet send = LongSet.create();
}else{
send.add(order.getId());
}
}
- collected in young generation

❖ V2: Preallocate all.
Order order = new Order();
buffer.fillOrder(order);
}else{
}
}
- one ‘static’ object per thread
- no allocations.

❖ V3: Access to serialised data instead object
class OrderAccess
{
static void getId(Buffer b,int offset)
{ return b.getLong(offset+0) }
static void putId(Buffer b, int offset, long id)
{ b.putLong(offset+0,id) }
static int getTimestamp(Buffer b, int offset)
{ return b.getInt(offset+8); }
……..
}

❖ V3: Access to serialised data instead object
for(int offset=0; !buffer.atEnd(offset);
offset+=OrderAccess.SIZE) {
long price = OrderAccess.getPrice(buffer,offset);
int quantity = OrderAccess.getQuantity(buffer,offset);
int fullPrice = price * quantity;
long partyLimit = OrderAccess.getPartyLimit(buffer,offset
if (partyLimit > maxLimit) {
client.rejected(buffer,offset,”party-limit”)
}else{
engine.process(buffer,offset);
}
}
- no temporary objects
- no allocations
- slightly unusual API

❖ V4: Offheap memory access via unsafe
sun.misc.Unsafe
❖ internal low-level sun API
❖ implemented by JIT intrinsic
❖ JVM crash on error
❖ widely used but not recommended
❖ will be deprecated in Java-9
❖ (removed in Java 10)
❖ (we will discuss alternatives)
class Unsafe
{
byte getByte(long addr)
void setByte(long addr, byte value)
….
byte getByte(Object value, long offs)
….
//same for primitive types.
….
// access to Object internals.
}

class OrderAccess
{
static void getId(long address)
{ return unsafe.getLong(address) }
static void putId(long address, long id)
{ unsafe.putLong(offset+0,id) }
static int getTimestamp(Buffer b, int offset)
{ return b.getInt(offset+8); }
……..
}

for(long address=buffer.begin; address!=buffer.end();
address=advance(address,OrderAccess.SIZE) ) {
long price = OrderAccess.getPrice(address);
int quantity = OrderAccess.getQuantity(address);
int fullPrice = price * quantity;
long partyLimit = OrderAccess.getPartyLimit(address)
if (partyLimit > maxLimit) {
client.rejected(address,”party-limit”)
}else{
engine.process(address);
}
}
- no allocations
- any memory
- subtle bugs are possible.
// better wrap by more hight-level interfaces (buffer, etc)

❖ V3, V4 Problem: Verbose API
- bytecode transformation from more convenient API
- Metaprogramming within hight-level JVM languages.
http://scala-miniboxing.org/ildl/
https://github.com/densh/scala-offheap
- Integration with unmanaged languages
- JVM with time-limited GC pauses. [Azul Systems]

Memory access API in next JVM versions
❖ VarHandlers (like methodHandlers)
❖ inject custom logic for variable access
❖ JEP 193 http://openjdk.java.net/jeps/193
❖ Value objects.
❖ can be located on stack
❖ JEP 169. http://openjdk.java.net/jeps/169
❖ Own intrinsics [project Panama]
❖ insert assembler instructions into JIT code.
❖ http://openjdk.java.net/projects/panama/
❖ JEP 191. http://openjdk.java.net/jeps/191
❖ ObjectLayout [Arrays 2.0]
❖ better contention
❖ http://objectlayout.github.io/ObjectLayout/

Low latency Java: access to native code
❖ JNA
❖ (idea — API to build native stack call and get result)
❖ pro: not require packaging native libraries.
❖ cons: SLOWWWW (but one-time: ok)
❖ JNI
❖ (idea — bridge between JNI/native code)
❖ pro: faster than JNA, exists critical jni
❖ cons: need packaging native library
❖ Shares Memory
❖ (idea — use same binary address from native & jvm code)
❖ pro: fast
❖ cons: unsafe, unusual API

Low latency Java: JNI
❖ JNI
❖ Java Object => Handle for native (maintain table)
❖ Callbacks are extremely slow
❖ Critical JNI
❖ not in ofﬁcial documentation
❖ substituted instead JNI after JI
// part of Cliff Click presentation
(advise for VM writers)

❖ JNI
❖ Java Object => Handle for native (maintain table)
❖ Callbacks are extremely slow
extern Java_nativeEventLoop(JEnv jenv, JObject job) {
while(!endOfBuffer(buff)) {
packet p = getData(buff);
if (checked(p)) {
processInJavaCallback(p);
}
}
}
C:
class JNIWrapper
{
void native eventLoop();
void processCallback(Packet p);
}
- p will be marshalled/ unmarshalled each time
- slow

extern Java_nativeEvent(JEnv jenv, JObject job) {
static int state = START;
switch(state) {
case START:
state = LOOP;
while(! endOfBuffer()) {
case LOOP:
Packet p = readData();
if (checked(p)) {
return (long)p;
}
}
}
return -1;
}
class JNIWrapper
{
void native event();
void process(long packetAddr
}
…
for(;;) {
p=event();
if (p==-1L) break;
process(p);
}
// better use return instead callback

❖ Critical JNI
❖ not in ofﬁcial documentation
❖ substituted instead JNI after JIT warm-up.
❖ only primitive types; static calls.
❖ must not be long-running. (GC is disabled)
extern long Java_nativeEvent(JEnv jenv, JObject job)
->
extern long JavaCritical_nativeEvent()
// want add timeout to satisfy ‘last’ condition

Low latency java: shared memory interface
❖ idea:
❖ spawn process or native thread.
❖ mmap file with given name to memory on ‘native
side’.
❖ mmap file with given name to memory on ‘java’ side.
❖ have shared memory persist to the same file in java &
native code.

Low latency Java/Native code data exchange
// buffer space
shared access
structure
struct BufferAccess
{
volatile long readIndex;
long[7] rpadding;
volatile long writeIndex;
long[7] wpadding;
}
C
class BufferAccessAccess // Java (example)
{
long getReadAddress(long baseAddr)
void setReadAddress(long baseAddr, long readAddr);
long getWriteAddress(long baseAddr)
}
base addr
Java

Low latency Java/fw
❖ Low-latency on JVM is
❖ possible
❖ sometimes non-trivial
❖ special techniques exists, in future will be better.

Thanks for attention.
❖ Questions ?
// ruslan@shevchenko.kiev.ua
// @rssh1
// https://github.com/rssh

Java & low latency applications

More Related Content

What's hot

Viewers also liked

Similar to Java & low latency applications

More from Ruslan Shevchenko

Recently uploaded

Java & low latency applications