combining levyx™ + intel optane™ to scale the …...combining levyx + intel® optane to scale...
TRANSCRIPT
Combining Levyx™ + Intel® Optane™ toScale the Performance of Matching Engines
LATENCY COMPARISON OF 1 BOOKMANAGER CORE, WITH PERSISTENCE (I.E. HELIUM) AND WITHOUT PERSISTENCE
Neglible Trade-off in Performance While Adding Critical Peristence Feature to the Working Data Set
USER SOFTWARE RUNNING ON
N CORE WITH SIMPLX
SCALABLE INTERFACE BETWEEN
SIMPLX AND HELIUM
HELIUM BUILT-IN MODE
THROUGHPUT (MSG/SEC)
LA
TE
NC
Y (
µse
c)
5,000 122,000 375,000
20
10
5
15
25
35
0
30
4099.9% WITH PERSISTENCE
99% WITH PERSISTENCE
99.9% WITHOUT PERSISTENCE
99% WITHOUT PERSISTENCE
SIMPLX CORE
Matching engines are at the core of electronic exchanges and use sophisticated algorithms to allocate trades among competing bids and offers at the same price. They match up bids and offers to complete trades. In addition to financial trading, matching engines and the need for them to process increasingly-large amounts of data are found in applications ranging from ad technology to interactive gaming to Smart Cities.
Simplx is a C++ development framework for building reliable cache-friendly distributed and concurrent multicore low-latency software. Levyx’s Helium™ engine combined with Intel® Optane™ create a solution that adds data persistence, a critical enterprise feature, without impacting the number of transactions that a financial bookmanager can process in a Simplx installation.
Note: Outlined functions are implemented in this demo.
• # of Books = 1000 => 1000 financial instruments
• Each BookManager handles many Books
• # of Book Managers is a scalability parameter
• Books are distributed evenly on the BookManager, hence on the cores
• Benchmarks run fully on one socket
• Each Persistence Core is linked to One BookManager Core
• 3 Injectors are used in this implementation
MATCHING ENGINES
THE PARAMETERS
THE IMPLEMENTATION
SIMPLX FRAMEWORK
One Book Core handles up to 125K orders/sec before performance degradation due to message queuing.
HELIUM API
SimplxTM
Simplx
Simplx
Simplx
HELIUM SERVER
OrderEntry Risk
Session management (open/close/auctions...)
+ Intel® OptaneTM DIMMs
Technical & business monitoring
Resilience & Persistence
...Book
ManagersMarket
DataPost
Trading
~10 usec differential in latency between5K and 122K msgs/sec is
inconsequential to most quants
Levyx was founded 2013 in Irvine California with a mission of enabling Real-Time Persistent Computing for Big Data™. To that end, Levyx has developed next-generation database Storage and Query offload engines that fully exploit the latest commodity hardware technologies including multi-core servers, internal and external flash systems, and IO offload engines.
The result is unprecedented performance and latency reductions for IO intensive workloads such financial service backtesting or streaming analytics.
Levyx’s new software-based engines allow for the first time, persistent computing to be possible on Big Data platforms such as Apache Spark thru use of SSDs instead of volatile memory-only designs.
Levyx is now delivering the world’s fastest key value store; the world’s first distributed storage/analytics offload engine (Xenon); and the world’s first large-scale, low-latency distributed database built for Exascale opportunities.
4 9 D i s c o v e r y , S u i t e # 2 2 0I r v i n e , C A 9 2 6 1 8
( 9 4 9 ) 5 0 2 - 6 3 6 9
s a l e s @ l e v y x . c o m
Trading Chain Scalability with Persistence and Latency Stability
Next-Generation Data Stack
Query/Analytics Parsing
Big Data Applications(Fintech, IoT, eCommerce, Genomics, ML/AI)
Query/Analytics Optimization & Partitioning Engine
Query/Analytics Code Generator & Just-in-time Compiler
Distributed Storage Class Memory (DSCM) Abstraction & API
NIC SSD SCM FLASH-ARRAY
Native Key/Value (NKV) Abstraction & API
Network & RDMA Manager (Synchroniza-
tion)
HighPerformance
Indexer
Lock Free Object Caching
& Write Buffering
Core Key/Value Store Logic
(PQ, RQ, GC & Transactions)
Key/Value Store API
DATA
STACK
UDF
GPU
Levyx Connectors
Ultra Low Latency - 10x Faster
Index Billions of Objectson a Single Node
Unmatched Price/Performance
ABOUT LEVYX
Note: Latency @ 99%ile, # of Book Cores vary from 1 to 7 (max on one socket)
Scaling the throughputwhile maintainingstable latency.
Dataset & Analytics API
TH
RO
UG
HP
UT
(K
OP
S/
SE
C)
LA
TE
NC
Y M
ICR
OS
EC
ON
DS
1-core 2-cores 4-cores 7-cores
300
100
0
500
700
400
200
600
800
900
6
2
0
10
14
8
4
12
16
18
122K ops/sec
230K ops/sec
500K ops/sec
800K ops/sec
11.5 µsec12.7 µsec
14.4 µsec
15.5 µsecLATENCY
THROUGHPUT
FPGA