Date post: | 10-May-2015 |
Category: |
Technology |
Upload: | peter-lawrey |
View: | 2,285 times |
Download: | 4 times |
GC Free coding
Peter LawreyCEO, Principal ConsultantHigher Frequency Trading
Agenda
Who are we?
When to optimise.
What can be achieved in extreme cases.
Examples of object replacement.
Libraries designed to be ultra-low GC.
Who are we
Higher Frequency Trading is a small consulting and software development house specialising in
Low latency, high throughput software 8 developers in Europe and USA. Sponsor HFT related open source projects Core Java engineering
Who am I?
Peter Lawrey
- CEO and Principal Consultant
- 3rd on Stackoverflow for Java, most Java Performance answers.
- Founder of the Performance Java User's Group
- An Australian, based in the U.K.
What is our OSS
Key OpenHFT projects Chronicle, low latency logging, event store and
IPC. (record / log everything) HugeCollections, cross process embedded
persisted data stores. (only need the latest)
Millions of operations per second.
Micro-second latency.
Why do we optimise for memory
The level required depends on the application Maximise throughput Remove worst pauses (Full GC, tenured) Remove small pauses Optimise CPU cache utilisation
Maximise Throughput
Latency is not important, only the amount of time it takes to perform a batch.
If you have a system which clean up do say 10 GB/s of garbage, it might be acceptable to spend 20% of your time GC-ing
In which case 2 GB/s of garbage might acceptable
Remove worst pauses
Stop the world, Full GCs and Tenure space collections.
An allocation rate of around 500 MB/s GC to below 5% of the CPU time reduce pressure on your tenured space Tuning the GC will do the rest
Remove small pauses
Reduce your garbage over a day to your Eden size.
Say you have a 24 GB Eden space and an allocation rate of 1 GB/hour, you can clean up once per day.
Remove small pauses
Time scales every developer should know.
Operation Latency In human terms
L1 Cache hit 1 ns A blink of an eye (~20 ms)
L2 Cache hit 3 ns Noticeable flicker
L3 Cache hit 10 – 20 ns Time to say “A”
Main memory 70 – 100 ns Time to say a ten word sentence
Signal down a 200m fibre cable
1 μsec One slide (speaking quickly)
SSD access 5 – 25 μsec Time to reheat a meal (3 mins)
HDD access 8 msec Time to flight around the world. (1.8 days)
Network packet from Germany to the USA
45 msec Waiting for a 7 working day delivery
Improve CPU cache utilisation
Your L1 cache is a small 32 KB of data, and L2 is 256 KB shared.
If you are producing 32 MB/s of garbage in a thread, you are filling your L1 cache with garbage every milli-seconds.
This makes it harder for the CPU to keep useful stuff in the cache. It doesn't know what is garbage.
Improving CPU Cache efficiency can speed up an application by 2-5x.
Why use Java?
A rule of thumb is that 90% of the time is spent in 10% of the code.
Writing in Java will mean that 10% of your code might mean optimising heavily.
Writing in C or C++ will mean that 100% of your code will be harder to write, or you have to use JNI, JNA, JNR-FFI.
Low level Java works well with natural Java.
When to optimise
Premature optimisation is the root of all evil
– Donald Knuth
Measure, Don't Guess
– Jack Shirazi & Kirk Pepperdine
When to optimise
You will need Key portion of code working correctly. A representative, reproducible test case. An idea of the optimisation requirements
Throughput or average latency. Latency profile at a given throughput.
Where to start
With representative and correct test case, I start with a commercial profiler with
CPU Profiling enabled Memory Profiling enabled.
With both on I look at the CPU profiling. Having the memory profiling on, gives memory allocation more weight.
An example: busy waiting for files
List<File> files = Arrays.asList(new File("dir/a.txt"), new File("dir/b.txt"), new File("dir/c.txt"));
// waiting for all files to existsboolean allFound;do { allFound = true; for (File file : files) { if (!file.exists()) { LockSupport.parkNanos(1); allFound = false; } }} while (!allFound);System.out.println("All files exist");
BTW DirectoryWatcher is the best solution.
Heap usage: SawTooth
Allocations
Green Level Optimisation
List<File> files = Arrays.asList(new File("dir/a.txt"), new File("dir/b.txt"), new File("dir/c.txt"));
// waiting for all files to existsboolean allFound;do { allFound = true; for (int i = 0; i < files.size(); i++) { if (!file.exists()) { LockSupport.parkNanos(1); allFound = false; } }} while (!allFound);System.out.println("All files exist");
Simple, saves 8% of garbage.
Red Level Optimisationstatic class MyFile extends File { String name; public MyFile(String pathname) { super(pathname); }
@Override public boolean exists() { return super.exists(); }
@Override public String getName() { if (name != null) return name; return name = super.getName(); }}public static void main(String... ignored) { List<MyFile> files = Arrays.asList(new MyFile("dir/a.txt"), new MyFile("dir/b.txt"), new MyFile("dir/c.txt"));
A bit scary and invasive but saves 46% of garbage.
Brown Level Optimisationprivate GetBytes getBytesCache;
public byte[] getBytes(String charsetName) throws UnsupportedEncodingException { if (charsetName == null) throw new NullPointerException(); GetBytes getBytes = getBytesCache; if (getBytes != null && getBytes.charsetName.equals(charsetName)) return getBytes.bytes; byte[] encode = StringCoding.encode(charsetName, value, 0, value.length); getBytesCache = new GetBytes(charsetName, encode); return encode;}
static class GetBytes { final String charsetName; final byte[] bytes; GetBytes(String charsetName, byte[] bytes) { this.charsetName = charsetName; this.bytes = bytes; }}
A very scary, uses -Xbootclasspath/p but saves 46% of garbage.
Heap usage: Flat line
Is it alive?
Common techniques to reduce garbage
Use primitives instead of wrappers.
Double → double
BigDecimal → double or long
Use primitive collections
Trove4j
Guava collections and utilities
Thread local mutable data
You can recycle data structures yourself if it is mutable.
Sharing mutable state between threads is hard.
Using mutable state which is local to a thread is much simpler and can be very low GC.
A quick win can be reusing buffers like ByteBuffer (and it can be off heap too)
SAX Parser instead of DOM
DOM is easier to work with, but requires reading in the whole document and building a tree of data.
SAX is harder to use, but you can build just the data structure you need as you go. You can populate a mutable structure for re-use.
Example, GC-free FIX parser
SAXophone is a library for SAX parsers FIX parser JSon parser BSon parser Yaml parser
SAX Parser interface
public interface BytesSaxParser { /** * reset any state */ void reset();
/** * Parse as much of the Bytes as possible. * * @param bytes */ void parse(Bytes bytes);}
FIX SAX Handler interface
public interface BytesSaxParser { /** * reset any state */ void reset();
/** * Parse as much of the Bytes as possible. * * @param bytes */ void parse(Bytes bytes);}
FIX Handlerfinal StringBuilder sender, target, clOrdId, symbol;double quantity, price;int ordType;
@Overridepublic void onField(long fieldNumber, Bytes value) { switch ((int) fieldNumber) { case 8: resetAll(); break; case 35: assert value.readByte() == 'D'; break; case 49: value.parseUTF(sender, StopCharTesters.ALL); break; case 56: value.parseUTF(target, StopCharTesters.ALL); break; case 11: value.parseUTF(clOrdId, StopCharTesters.ALL); break; case 55: value.parseUTF(symbol, StopCharTesters.ALL); break; case 38: quantity = value.parseLong(); break; case 44: price = value.parseDouble(); break; case 40: ordType = (int) value.parseLong(); break; case 10: processFixMessage(sender, target, clOrdId, symbol, quantity, price, ordType); break; } count.incrementAndGet();}
FIX Message to Decode
String s = "8=FIX.4.2|9=130|35=D|34=659|49=BROKER04|56=REUTERS|52=20070123-19:09:43|38=1000|59=1|100=N|40=1|11=ORD10001|60=20070123-19:01:17|55=HPQ|54=1|21=2|10=004|";
NativeBytes nb = new DirectStore(s.length()).bytes();
nb.append(s.replace('|', '\u0001'));
nb.flip();
Test harness
final AtomicInteger count = new AtomicInteger();FixSaxParser parser = new FixSaxParser(new MyFixHandler(count));
int runs = 200000;for (int t = 0; t < 5; t++) { count.set(0); long start = System.nanoTime(); for (int i = 0; i < runs; i++) { nb.position(0); parser.parse(nb); } long time = System.nanoTime() - start; System.out.printf("Average parse time was %.2f us, " + "fields per message %.2f%n", time / runs / 1e3, (double) count.get() / runs);}
Note : this is average latency which is really a proxy for throughput. Average latency hides large delays, but in the sub-microsecond range it is harder to get a meaningful latency distribution.
PerformanceRun with -verbose:gc -Xmx32m
Average parse time was 0.96 us, fields per message 17.00
Average parse time was 0.58 us, fields per message 17.00
Average parse time was 0.58 us, fields per message 17.00
Average parse time was 0.55 us, fields per message 17.00
Average parse time was 0.54 us, fields per message 17.00
Q & A
https://github.com/OpenHFT/OpenHFT
@PeterLawrey
What is HFT?
No standard definition. Trading faster than a human can see. Being fast can make the difference between
making and losing money. For different systems this means typical
latencies of between 10 micro-seconds and 10 milli-second.
(Latencies external to the provider)
Time scales every developer should know.
Operation Latency In human terms
L1 Cache hit 1 ns A blink of an eye (~20 ms)
L2 Cache hit 3 ns Noticeable flicker
L3 Cache hit 10 – 20 ns Time to say “A”
Main memory 70 – 100 ns Time to say a ten word sentence
Signal down a 200m fibre cable
1 μsec One slide (speaking quickly)
SSD access 5 – 25 μsec Time to reheat a meal (3 mins)
HDD access 8 msec Time to flight around the world. (1.8 days)
Network packet from Germany to the USA
45 msec Waiting for a 7 working day delivery
Simple Trading System
Event driven processing
Trading system use event driven processing to minimise latency in a system.
Any data needed should already be loaded in memory, not go off to a slow SQL database.
Each input event triggers a response, unless there is a need to limit the output.
Critical Path
A trading system is designed around the critical path. This has to be as short in terms of latency as possible.
Critical path has a tight latency budget which excludes many traditional databases.
Even the number of network hops can be minimised.
Non critical path can use tradition databases
Critical Path databases
Time Series databases Kdb, kona InfluxDB OpenTSDB
Designed for millions of writes per second.
Column based database => 100 Million operations per second e.g. sum a column.
Critical Path Databases
Critical Path data store
HFT strategies are; described using graphs. handle events in real time ~10 – 100 μsec. cache state rather than query a database. all custom written libraries AFAIK.
Critical Path data store
Logging is performed by appending to memory mapped files.
OpenHFT's Java Chronicle makes this easier to do in Java in a GC-free, off heap, lock less way.
Such low level coding is relatively easy in C or C++.
Non-critical Datastore
Configuration management ZooKeeper, etcd Plain files with Version control LDAP Any distributed key-value store. e.g. MongoDB
Big Data
Back testing a HFT system is critical and a number of solutions are available
Hadoop Matlab Time series R
Operational Infrastructure
Control and management infrastructure JMS, JMX Tibco RV, LBM Terracotta MongoDB
Reliable persistence
Trades and Orders are high value data and less voluminous than Market data or strategy results.
Typically SQL Database. Sometimes multiple databases for different
applications.
Why use more exotic database?
Mostly for high throughput. Million per second in one node.
Often for low latency. Latencies well below a milli-second.
Why wouldn't you use exotic DB
Not easy to learn, high knowledge investment.
(!R)@&{&/x!/:2_!x}'!R Often harder to use.
Less management tools. Not designed to work with web applications.
More sensitive to the details of the hardware and what else is running on the same machine.
Low latency at high throughput
Java Chronicle is designed as a low latency logger and IPC.
At one million small messages per second Almost zero garbage Latency between processes around 1 micro-
second Concurrent readers and writers
Supports bursts of 10 million messages/sec.
Chronicle and replication
Replication is point to point (TCP)
Server A records an event
– replicates to Server B
Server B reads local copy
– B processes the event
Server B stores the result.
– replicates to Server A
Server A replies.
Round trip 25 micro-seconds99% of the time
GC-freeLock lessOff heap
Unbounded
HugeCollections
HugeCollections provides key-value storage. Persisted (by the OS) Embedded in multiple processes Concurrent reads and writes Off heap accessible without serialization.
HugeCollections and throughput
SharedHashMap tested on a machine with 128 GB, 16 cores, 32 threads.
String keys, 64-bit long values. 10 million key-values updated at 37 M/s 500 million key-values updated at 23 M/s On tmpfs, 2.5 billion key-values at 26 M/s
HugeCollections and latency
For a Map of small key-values (both 64-bit longs)
With an update rate of 1 M/s, one thread.
Percentile 100K
entries1 M entries 10 M entries
50% (typical) 0.1 μsec 0.2 μsec 0.2 μsec
90% (worst 1 in 10) 0.4 μsec 0.5 μsec 0.5 μsec
99% (worst 1 in 100) 4.4 μsec 5.5 μsec 7 μsec
99.9% 9 μsec 10 μsec 10 μsec
99.99% 10 μsec 12 μsec 13 μsec
worst 24 μsec 29 μsec 26 μsec
Bonus topic: Units
A peak times an application writes 49 “mb/s” to a disk which supports 50 “mb/s” and is replicated over a 100 “mb/s” network.
What units were probably intended and where would you expect buffering if any?
Bonus topic: Units
A peak times an application writes 49 MiB/s to a disk which supports 50 MB/s and is replicated over a 100 Mb/s network.
MiB = 1024^2 bytes
MB = 1000^2 bytes
Mb = 125,000 bytes
The 49 MiB/s is the highest rate and 100 Mb/s is the lowest.
Bonus topic: Units
Unit bandwidth Used for
mb - miili-bit mb/s – milli-bits per second ?
mB - milli-byte mB/s – milli-bytes per second ?
kb – kilo-bit (1000) kb/s – kilo-bits (baud) per second Dial up bandwidth
kB – kilo-byte (1000) kB/s – kilo-bytes per second ?
Mb – mega-bit (1000^2) Mb/s – mega-bits (baud) per second Cat 5 ethernet
MB - mega-byte (1000^2) MB/s – mega bytes per second Disk bandwidth
Mib – mibi-bit (1024^2) Mib – Mibi-bits per second ?
MiB – mibi-byte (1024^2) MiB – Mibi-bytes per second Memory bandwidth
Gb – giga-bit (1000^3) Gb/s – giga-bit (baud) per second High speed networks
GB – giga-byte (1000^3) GB/s – giga-byte per second -
Gib – gibi-bit (1024^3) Gib/s – gibi-bit per second -
GiB – gibi-byte (1024^3) GiB/s – gibi-byte per second. Memory Bandwidth