IMC Summit 2016 Breakout - William Bain - Implementing Extensible Data Structures in In-Memory Data Grids

IMPLEMENTING EXTENSIBLE DATA
STRUCTURES IN IN-MEMORY DATA GRIDS
WILLIAM L. BAIN
See all the presentations from the In-Memory Computing
Summit at http://imcsummit.org

AGENDA
 The Need for IMDG-Based Data Structures:
 Brief Review of Using In-Memory Data Grids (IMDGs) for Data Access & Query
 Computing in the Client: the Network Bottleneck
 Build Data Structures in the Grid:
 Using the Grid as an Extensible Data-Structure Store (e.g., Redis)
 Challenges with Running in the Grid Service
 Build Extensible Data Structures Outside the Grid Service:
 Example of “Single Method Invocation” (SMI)
 The Next Step: Data-Parallel Data Structures and Methods:
 Example of “Parallel Method Invocation” (PMI)
 Sharding Data Structures for Data-Parallelism
 Combining Single and Data-Parallel Operations
2

 Dr. William Bain, Founder & CEO of ScaleOut Software:
 Ph.D. in Electrical Engineering (Rice University, 1978)
 Career focused on parallel computing – Bell Labs, Intel, Microsoft
 3 prior start-ups, last acquired by Microsoft and product now ships as Network Load Balancing in Windows Server
 ScaleOut Software develops and markets In-Memory Data Grids,
software middleware for:
 Scaling application performance and
 Providing operational intelligence using
 In-memory data storage and computing
 Eleven years in the market; 425+ customers, 10,000+ servers
ABOUT THE SPEAKER

THE NEED FOR IMDG-
BASED DATA
STRUCTURES

WHAT ARE IN-MEMORY DATA GRIDS?
 In-memory data grid (IMDG) provides scalable, hi av storage for live data:
 Designed to manage business logic state:
 Sequentially consistent data shared by multiple clients
 Object-oriented collections by type
 Create/read/update/delete APIs for Java/C#/C++
 Parallel query by object properties
 Designed for transparent scalability and high availability:
 Automatic load-balancing across commodity servers
 Automatic data replication, failure detection, and
recovery
 IMDG also provides an ideal platform for “operational
intelligence”:
 Easy to track live systems with large workloads
 Appropriate availability model for production deployments
5

COMPARING IMDGS TO SPARK
6
IMDGs Spark
Data
characteristics
“Live” key-value
pairs
Static data or batched
streams
Data model for
APIs
Object collections Resilient distributed
datasets
Focus of APIs CRUD, query, data-
structure methods
Data-parallel operators on
RDDs for analytics
High availability
tradeoffs
Data replication for
fast recovery
Lineage for max
performance
 On the surface, both are surprisingly similar:
 Both designed as scalable, in-memory computing platforms
 Both implement data-parallel operators
 Both can handle streaming data
 But there are key differences that impact use in live systems:

IMDG’S DATA ACCESS MODEL
 Object-oriented APIs store unstructured collections of objects by
type:
 Accessible by key (string or number) using “CRUD” APIs
 APIs in Java, C#, C++, REST, …
 Also accessible by distributed query on properties
 Stored within IMDG as serialized blobs or data structures
 Objects often have extended semantics:
 Synchronization with distributed locking
 Bulk loading
 Timeouts, eviction
 Transparent access to a backing store or remote IMDG
 Methods on data structures
7
Basic “CRUD” APIs:
• Create(key, obj, tout)
• Read(key)
• Update(key, obj)
• Delete(key)
and…
• Lock(key)
• Unlock(key)
Object
key

EXAMPLE OF JAVA ACCESS APIS
8
static void Main(string argv[])
{
// Initialize string object to be stored:
String s = “Test string”;
// Create a cache collection:
SossCache cache = SossCacheFactory.getCache(“MyCache”);
// Store object in the grid:
CachedObjectId id = new CachedObjectId(UUID.randomUUID());
cache.put(id, s);
// Read object stored in the grid:
String answerJNC = (String)cache.get(id);
// Remove object from the grid:
cache.remove(id);}

EXAMPLE OF C# ACCESS APIS
9
static void Main(string[] args)
{
// Initialize object to be stored:
SampleClass sampleObj = new SampleClass();
sampleObj.var1 = "Hello, grid!";
// Create a cache:
SossCache cache = CacheFactory.GetCache("myCache");
// Store object in the grid:
cache["myObj"] = sampleObj;
// Read object stored in grid:
retrievedObj = cache["myObj"] as SampleClass;
// Remove object from the grid:
cache["myObj"] = null;
}

TRANSPARENT ACCESS TO REMOTE IMDGS
 Use CRUD APIs to
transparently access
objects held within
remote IMDGs.
 Implements a virtual
IMDG spanning
multiple sites.
 IMDG can implement
coherency policies to
mitigate WAN latency.
10
Remote Site A Remote Site B
READ
WAN WAN
Local Site
Globally Distributed IMDG

USING PARALLEL QUERY TO ACCESS DATA
 Retrieves multiple objects with selected properties
and/or tags.
 Runs on all grid servers to scale query performance.
 Typically uses fast, indexed lookup on each grid server.
 Often used to read a set of related objects for analysis.
 However, has O(N) overhead per query:
 Implies overall O(N**2) overhead with many clients
 Can easily lead to network saturation
 Not suitable as primary access method
11
Queries the
IMDG
in parallel.
Merges the
keys
into a list for the
client.

PARALLEL QUERY EXAMPLE (JAVA)
 Mark selected class properties as indexes for parallel query:
 Define query using these properties:
12
public class Stock implements Serializable {
private String ticker;
private int totalShares;
private double price;
@SossIndexAttribute
public String getTicker() {return ticker;} … }
NamedCache cache = CacheFactory.getCache("Stocks",
false);
Set keys = cache.queryKeys(Stock.class,
or(equal("ticker", "GOOG"),
equal("ticker", "ORCL")));

PARALLEL QUERY EXAMPLE (C# WITH LINQ)
 Mark selected class properties as indexes for parallel query:
 Define query using these properties; optionally access objects automatically
from IMDG:
13
class Stock {
[SossIndex]
public string Ticker { get; set; }
public decimal TotalShares { get; set; }
public decimal Price { get; set; }}
NamedCache cache = CacheFactory.GetCache("Stocks");
var q = from s in cache.QueryObjects<Stock>()
where s.Ticker == "GOOG" || s.Ticker == "ORCL"
select s;
Console.WriteLine("{0} Stocks found", q.Count());

RUNNING A METHOD IN THE CLIENT
14
 Client reads or queries objects, run a method, and then optionally update the grid.
 Data motion can place a huge burden on the network.
 This approach fails to make use of untapped
computing power in IMDG hosts.

Example: E-commerce inventory management:
• E-commerce site stores orders and inventory changes within IMDG.
• Goal: reconcile orders and inventory by SKU in real time for 50K SKUs.
• Implementation using parallel query (pseudo-code):
• Problem: Queries quickly saturate the network.
Multiple clients
QUERY CAN CAUSE NETWORK SATURATION
15
parallel foreach (sku){
orders[] = QueryIMDG(orders_ns, sku);
stockch[] = QueryIMDG(stock_ns, sku);
Reconcile(orders, stockch);
SendAlerts();}

THE EFFECT OF DATA MOTION
 Network access creates a
bottleneck that limits
throughput and increases
latency.
 Avoiding data motion
enables linear scalability for
growing workloads =>
predictable, low latency.
 Example: back-testing stock
histories in parallel
 How? Move the
computation
into the grid to avoid data
motion.
16

BUILDING DATA
STRUCTURES IN THE
GRID SERVICE

 Client invokes method and ships parameters to an IMDG object.
 Parameters are typically much smaller than the object.
 IMDG runs the method in the local grid service and
returns a result to the client.
 Data motion is dramatically reduced.
RUNNING METHODS IN AN IN-MEMORY DATA GRID
18

 IMDG becomes a data structure store that implements useful data types and methods.
 Example: Redis (which provides lists, sets of strings, hashed sets, sorted sets)
 Advantages: fast (reduced data motion, no deserialization),
secure, less client development
BUILDING PREDEFINED DATA STRUCTURES IN THE GRID
19

 Stores an ordered sequence of string values within a single Redis object.
 Implements client APIs for managing the list.
 Allows Redis server to move
items between lists.
 Client avoids reading and updating
entire list object to perform list
operations.
SIMPLE EXAMPLE: REDIS LIST
20
LIST:
RPUSH key-name value [value]
LPUSH key-name value [value]
RPOP key-name returns value
LPOP key-name returns value
LRANGE key-name start end
LTRIM key-name start end
BLPOP key-name [key-name …]
BRPOP key-name [key-name …]
RPOPLPUSH src-key dst-key
BRPOPLPUSH src-key dst-key

BUILDING EXTENSIBLE DATA STRUCTURES IN REDIS
Users can extend data structures using
Lua server-side scripts:
 Operate on local server objects.
 Run without interruption to implement
atomic operations (avoiding multiple round
trips).
 Call standard Redis commands within the
script.
 Use to extend existing data structures or
implement new ones (e.g., sharded list).
Breaking news: Upcoming Redis release will
have loadable modules for greater
extensibility.
21
>>> r = redis.StrictRedis()
>>> lua = """
... local value = redis.call('GET', KEYS[1])
... value = tonumber(value)
... return value * ARGV[1]"""
>>> multiply = r.register_script(lua)
>>> r.set('foo', 2)
>>> multiply(keys=['foo'],args=[5])
10
Example of using a Lua script (Python)
from https://pypi.python.org/pypi/redis

ISSUES WITH RUNNING IN THE GRID SERVICE
 Performance:
 Running complex methods can slow down grid service’s data access.
 A hot object can consume CPU and delay other accesses and methods.
 Extensibility:
 Local server scripting does not support distributed data structures.
 Scripting engine must be compatible with service’s language runtime environment.
 Reliability/Security:
 Extensible server-based scripting can crash or compromise the server.
 Alternative: run data structure methods in a separate worker process.
 Solves extensibility issues: security, language flexibility, distributed access.
 But…creates tradeoffs: performance, complexity.
22

BUILDING EXTENSIBLE
DATA STRUCTURES
OUTSIDE THE GRID
SERVICE

USING AN INVOCATION GRID FOR DATA STRUCTURES
 On each IMDG host, create a
worker process which executes
data-structure methods.
 The set of worker processes is
called an invocation grid.
 IGs usually run language-specific
execution environments (JVM,
.NET runtime).
 IGs handle execution requests
from the local grid service.
 Data-structure methods can
access all objects in the IMDG.
24
Invocation Grid

RUNNING USER CODE IN AN INVOCATION GRID
25
Execution steps for running a method:
 Client invokes method on a specific object.
 Client library maps request to a grid host.
 Grid service forwards request to local worker
process.
 Worker process executes method:
 Grid service forwards serialized object and
parameters to the worker; to mitigate overhead:
 Worker maintains a client-side cache of recently accessed
objects.
 Worker can use memory-mapped file.
 Method optionally returns result to grid service.
 Grid service optionally returns result to client.

ONE TECHNIQUE: “SINGLE METHOD INVOCATION”
Leverages Object-Oriented View of Grid Data:
 Assume that an IMDG collection (“name space”) contains objects of a single type.
 Each object encapsulates a data structure managed by user-defined code.
 SMI invokes user-defined method on a specified instance within the name space using object’s key:
 Optionally passes in a parameter object.
 Optionally returns a result object.
 Can encapsulate SMI call into user-defined API wrapper to simplify client’s view:
ResultType ClientMethod(GridKey key, ParamType param1, ...) {
<setup namespace>;
<copy parameters into params_obj>;
return namespace.SingleObjectInvoke(key, SmiMethod, param_obj);}
result = namespace.SingleObjectInvoke(key, SmiMethod, param_obj);
26

EXAMPLE: REDIS LIST DATA STRUCTURE USING SMI (C#)
 Create server-side data structure and methods:
 Run client wrapper which invokes server-side method on target host (pseudo-code here):
27
public class StackList<T> : LinkedList<T> {
public StackList() : base() {}
public static void LPush(StackList<T> list, T item) {list.AddFirst(item);}
public static void RPush(StackList<T> list, T item) {list.AddLast(item);}
public static T LPop(StackList<T> list) {
T retItem = (list.First as LinkedListNode<T>).Value;
base.RemoveFirst(); return retItem;}
public static T RPop(StackList<T> list) {
T retItem = (list.Last as LinkedListNode<T>).Value;
base.RemoveLast(); return retItem;}}
InvokeLPush<T>(GridKey list-key, LinkedListNode<T> item) {
return nc.SingleObjectInvoke(list-key, LPush, item);}

EXAMPLE: USER-DEFINED DATA STRUCTURE (JAVA)
28
public class StockPriceData {
private double stockPrice;
private double peRatio; … };
StockPriceData[] data = new StockPriceData[]{
new StockPriceData(100.3, 10.4),
new StockPriceData(102.1, 20.1), …};
NamedCache pset = CacheFactory.getCache("stockCache");
pset.add(“stockData", data);
public class StockInvokable implements Invokable<StockPriceData[],
Integer, Double> {
public Double eval(StockPriceData[] data, Integer param){
<analyze data and return Double result> }};
Double result = pset.singleObjectInvoke(new StockInvokable(),“stockData",data.length);

 First obtain a reference to the IMDG’s object collection of portfolios:
 First obtain a reference to the IMDG’s object collection of portfolios:
 Create an “invocation grid,” a re-usable compute engine for the application:
 Spawns a JVM on all grid servers and connects them to the in-memory data grid.
 Stages the application code and dependencies on all JVMs.
 Associates the invocation grid with an object collection.
29
InvocationGrid grid = new InvocationGridBuilder("grid")
.addClass(DependencyClass.class)
.addJar("/path/to/dependency.jar")
.setJVMParameters("-Xmx2m")
.load();
pset.setInvocationGrid(grid);
NamedCache pset = CacheFactory.getCache(“portfolios");
EXPLICITLY CREATING AN INVOCATION GRID

30
IMDG CREATES IG WORKER PROCESSES IN PARALLEL

RUNNING MULTIPLE REQUESTS IN PARALLEL
31
Concurrent method execution provides parallel
speedup:
• IMDG handles multiple streaming requests
from a single client.
• Also handles multiple clients in parallel.
• Provides “embarrassingly parallel”
speedup.

PROS AND CONS OF USING AN INVOCATION GRID
Advantages:
 Secure:
 User code cannot access grid data.
 User code cannot bring down grid service.
 Flexible:
 Supports arbitrary languages.
 Supports full IMDG APIs and access to remote objects.
 Enables building distributed data structures.
Disadvantages:
 Higher latency due to round trip from grid service to IG worker:
 Mitigated by client cache in IG worker and memory-mapped files
 Additional CPU and data motion on grid hosts
 Complexity of implementing code shipping and IG worker orchestration
32

THE NEXT STEP: DATA-
PARALLEL DATA
STRUCTURES AND
METHODS

WHY USE DATA-PARALLEL METHODS?
Leverages the grid’s scalability to keep latency low for large workloads.
 Limitations of single method execution:
 Data structures must fit within a single object.
 Does not leverage scalability of the IMDG to
perform a single operation on many objects.
 Data-parallel method execution:
 Implements a data structure operation spanning
multiple objects.
 Examples: MapReduce, weather simulation
 Runs computation in parallel on all objects and
optionally combines results.
 Benefit: handles very large data structures with low
latency and scalable speedup.
34
Invoke Method
(Eval)
Combine Results
(Merge)

DATA-PARALLEL METHOD EXECUTION
35
 Client runs a method on multiple objects
distributed across all grid hosts.
 Objects are usually part of one typed
collection.
 Results are merged into a single result
returned to the client.
 Provides a building block for complex
computations (e.g., MapReduce).

Like SMI, PMI leverages object-oriented view of grid data:
 Follows standard HPC model (broadcast/merge) for data-parallel computation.
 Assumes that an IMDG collection (name space) contains objects of a single type.
 Each object maintains a data structure managed by user-defined code.
 PMI invokes user-defined “eval” method in parallel on a set of queried instances within the name
space:
 Optionally passes in a parameter object.
 Optionally returns a result object.
 PMI invokes optional user-defined “merge” method in parallel to perform binary merging across all
results:
 Returns a single result to the client (or serves as a barrier).
 Uses binary merge tree to combine all results in log-2(N) time where N is # grid hosts.
36
result = namespace.ParallelInvoke(query_spec, Eval, Merge, param_obj);
ONE TECHNIQUE: “PARALLEL METHOD INVOCATION”

EXAMPLE IN FINANCIAL SERVICES
 Goal: track market price fluctuations for a hedge fund and
keep portfolios in balance.
 How:
 Keep portfolios of stocks (long and short positions)
in object collection within IMDG.
 Collect market price changes in
one-second snapshots.
 Define a method which applies a
snapshot to a portfolio and optionally
generates an alert to rebalance.
 Perform repeated parallel method invocations
on a selected (i.e., queried) set of portfolios.
 Combine alerts in parallel using a user-defined merge
method.
 Report alerts to UI every second for fund manager.
37

UI FOR FINANCIAL SERVICES EXAMPLE
 UI performs a PMI every
second with a new market
snapshot.
 User can select filters to
parameterize PMI.
 PMI completes in ~350 msec.
 UI lists all portfolios (on left).
 UI highlights alerted portfolios
from PMI merge.
 User selects and reads portfolio
object within grid to see alerted
positions (on right).

PORTFOLIO EXAMPLE: DEFINING THE DATASET
 Simplified example of a portfolio class (Java):
 Note: some properties are made query-able.
 Note: the evalPositions method analyzes the portfolio for a market snapshot.
39
public class Portfolio {
private long id;
private Set<Stock> longPositions;
private Set<Stock> shortPositions;
private double totalValue;
private Region region;
private boolean alerted; // alert for trading
@SossIndexAttribute // query-able property
public double getTotalValue() {…}
@SossIndexAttribute // query-able property
public Region getRegion() {…}
public Set<Long> evalPositions(MarketSnapshot ms) {…};}

DEFINING THE PARALLEL METHODS
 Implement interface that defines methods for analyzing each object and for merging the results:
40
public class PortfolioAnalysis implements
Invokable<Portfolio, MarketSnapshot, Set<Long>>
{
public Set<Long> eval(Portfolio p, MarketSnapshot ms)
throws InvokeException {
// update portfolio and return id if alerted:
return p.evalPositions(ms);
}
public Set<Long> merge(Set<Long> set1, Set<Long> set2)
throws InvokeException {
set1.addAll(set2);
return set1; // merged set of alerted portfolio ids
}}

INVOKING THE PARALLEL METHOD INVOCATION
 Run the PMI on a queried set of objects within the collection:
 Multicasts the invocation and parameters to all JVMs.
 Runs the data-parallel computation.
 Merges the results and returns a final result to the point of call.
41
InvokeResult alertedPortolios = pset.invoke(
PortfolioAnalysis.class,
Portfolio.class,
and(greaterThan(“totalValue”, 1000000), // query spec
equals(“region”, Region.US)),
marketSnapshot, // parameters
...);
System.out.println("The alerted portfolios are" +
alertedPortfolios.getResult());

EXECUTION STEPS
 Eval phase: each server queries local objects and
runs eval and merge methods:
 Note: accessing local data avoids networking
overhead.
 Completes with one result object per server.
42
 Merge phase: all servers perform distributed
merge to create final result:
 Merge runs in parallel to minimize completion time.
 Returns final result object to client.

THE GOAL: SCALABLE SPEEDUP
 Assume that a workload of size W requires time T to complete on a single
server:
 Throughput on 1 server = 1/T
 Latency on 1 server = T
 Non-goal: Use N servers to decrease latency to T/N for fixed-size workload
W.
 Goal: linearly scale throughput while workload grows:
 Workload grows N-fold on N servers.
 Throughput on N servers = N/T for workload of size N*W.
 Latency on N servers = T (does not grow).
 Example: web farm from handling W users to N*W users.
 Bottlenecks to speedup lower throughput and increase latency:
 Usually due to network congestion.
 Can be due to other resource contention (e.g., hot objects).
add servers
grow workload
latency

IMDG DELIVERS SCALABLE SPEEDUP
Example: Measured a similar financial services application (back testing stock trading strategies on stock
histories):
 Hosted IMDG in Amazon EC2 using 75
servers holding 1 TB of stock history
data in memory
 IMDG handled a continuous
stream of updates (1.1 GB/s)
 Results: analyzed 1 TB in
4.1 seconds (250 GB/s).
 Observed near linear scaling as
dataset and update rate grew.
44

SCALING OPERATIONS ON DATA STRUCTURES
Sharding allows concurrent (and data-parallel) operations and enables scalable speedup.
 Step 1: Sharding within a single server:
 Splits a single object into multiple parts (“shards”).
 Allows concurrent operations on different
shards.
 Takes advantage of multiple CPU cores
within a server for parallel speedup.
 Example: sharding a Redis list (from Redis in Action)
 Uses Lua server-side script.
 Splits list into shard objects.
 Uses well-known local objects to find list ends.
 Limitation: cannot distribute shards across servers.
45
list:5 list:6 list:7list:first
list:5
list:last
list:7

SCALING ACROSS MULTIPLE SERVERS
 Step 2: Sharding across multiple servers:
 Handles N-times larger workload (memory use) on N servers.
 Enables linearly increasing throughput and maintains low latency as workload grows.
 One technique: shard in client:
 Clients map shards to servers.
 Servers manage shards
independently
(“embarrassingly parallel”).
 Example: distributed Redis
hashed set
 Issue: how to re-shard
when servers are added or
removed?
46

EXAMPLE: JAVA CONCURRENT MAP
 Challenge: How to store very large number of small
key-value pairs within an IMDG:
 Access KVPs individually using Java concurrent map
semantics.
 Access/update KVPs in parallel using a MapReduce
engine for analysis.
 One implementation (ScaleOut “Named Map”):
 Shard chunks of KVPs across IMDG servers.
 Use hash table within each chunk for fast lookup.
 Pipeline chunks to/from MapReduce engine as
splits/partitions.
 Host chunks as normal IMDG objects (with server-based
sharding) for automatic load-balancing.
47

COMBINING SINGLE
AND DATA-PARALLEL
OPERATIONS ON DATA
STRUCTURES

COMBINING SMI & PMI IN ONE APPLICATION
 Example: Tracking cable TV set-top boxes
 Goals:
 Make real-time, personalized upsell offers
 Detect and manage service issues
 Track aggregate behavior to identify patterns, e.g.:
 Total instantaneous incoming event rate, network hot spots
 Most popular programs and # viewers by zip code
 How:
 Represent set-top-boxes as data structure objects held in the IMDG.
 Use SMI to track events from 10M set-top boxes with 25K events/sec (2.2B/day):
 Correlate, cleanse, and enrich events per rules (e.g. ignore fast channel switches, match channels to programs)
 Feed enriched events to recommendation engine within 5 seconds
 Use PMI with a named map to track aggregate statistics across all set-top boxes every 10 seconds.
49
©2011 Tammy Bruce presents LiveWire

 Each set-top box is represented as an
object in the in-memory data grid (IMDG).
 Object holds raw & enriched event
streams, viewer parameters, and statistics
 IMDG captures incoming events and
correlates/cleanses/enriches using SMI:
 Note: IMDG makes an excellent stream-
processing engine.
 Avoids data motion of other approaches.
 IMDG uses PMI across all set-top box
objects to periodically collect and report
global statistics.
USING AN IMDG TO TRACK SET-TOP BOXES
50

PERFORMING CROSS-SERVER SMI
 SMI may need to invoke a second method
on a remote object.
 Example: enriching channel-change events
from a program guide
 Requirements:
 SMI must have global view of IMDG objects.
 SMI must be able to allow asynchronous
execution of a remote request.
 Solution:
 Run method within IG worker (vs. grid
service).
 Enable fully distributed view of IMDG in SMI.
51

Example from a proof of concept UI based on a
simulated workload for San Diego metropolitan area:
 Runs on a ten-server cluster.
 Continuously correlates and cleanses telemetry from
10M simulated set-top boxes (from synthetic load
generator).
 Processes more than 30K events/second.
 Enriches events with program information every
second.
 Tracks and reports aggregate statistics (e.g., top 10
programs by zip code) every 10 seconds.
THE RESULT: REAL-TIME STREAMING & STATISTICS
Real-Time Dashboard
52

RECAP: ADDING DATA STRUCTURES TO IMDGS
 Built-in data structures (e.g., Redis) provide a pre-defined set of fast,
easy to use data structures.
 Moving computation into an in-memory data grid offers several
benefits:
 Reduces network bottlenecks due to data motion.
 Leverages CPU power of the IMDG’s hosts.
 Enables scalable speedup to handle large workloads with low latency.
 Use of worker process increases security, reliability, and flexibility,
although it increases overhead.
 Extensible, grid-based data structures enable efficient
implementation
of complex computations.
 These techniques can be used for both single-object and data-parallel
computations.
Bottom line: IMDGs are a great platform for in-memory computing.
53

IMC Summit 2016 Breakout - William Bain - Implementing Extensible Data Structures in In-Memory Data Grids

More Related Content

What's hot (20)

Viewers also liked (11)

Similar to IMC Summit 2016 Breakout - William Bain - Implementing Extensible Data Structures in In-Memory Data Grids (20)

More from In-Memory Computing Summit (20)

Recently uploaded (20)

IMC Summit 2016 Breakout - William Bain - Implementing Extensible Data Structures in In-Memory Data Grids