Tuesday, 30 March 2010

Infinispan eviction, batching updates and LIRS

DataContainer abstraction represents the heart of Infinispan. It is a container structure where actual cache data resides. Every put, remove, get and other invoked cache operations eventually end up in the data container. Therefore, it is of utmost importance the data container is implemented in a way that does not impede overall system throughput. Also recall that the data container’s memory footprint can not grow indefinitely because we would eventually run out of memory; we have to periodically evict certain entries from the data container according to a chosen eviction algorithm.

LRU eviction algorithm, although simple and easy to understand, under performs in cases of weak access locality (one time access entries are not timely replaced, entries to be accessed soonest are unfortunately replaced, and so on). Recently, a new eviction algorithm - LIRS has gathered a lot of attention because it addresses weak access locality shortcomings of LRU yet it retains LRU’s simplicity.

However, no matter what eviction algorithm is utilized, if eviction is not implemented in a scalable, low lock contention approach, it can seriously degrade overall system performance. In order to do any meaningful selection of entries for eviction we have to lock data container until appropriate eviction entries are selected. Having such a lock protected data container in turn causes high lock contention offsetting any eviction precision gained by sophisticated eviction algorithms. In order to get superior throughput while retaining high eviction precision we need both low lock contention data container and high precision eviction algorithm implementation – a seemingly impossible feat.

Instead of making a trade-off between the high precision eviction algorithm and the low lock contention there is a third approach: we keep lock protected data container but we amortize locking cost through batching updates. The basic idea is to wrap any eviction algorithm with a framework that keeps track of cache access per thread (i.e. ThreadLocal) in a simple queue. For each cache hit associated with a thread, the access is recorded in the thread’s queue. If thread’s queue is full or the number of accesses recorded in the queue reaches a certain pre-determined threshold, we acquire a lock and then execute operations defined by the eviction algorithm - once for all the accesses in the queue. A thread is allowed to access many cache items without requesting a lock to run the eviction replacement algorithm, or without paying the lock acquisition cost. We fully exploit a non-blocking lock APIs like tryLock. As you recall tryLock makes an attempt to get the lock and if the lock is currently held by another thread, it fails without blocking its caller thread. Although tryLock is cheap it is not used for every cache access for obvious reasons but rather on certain pre-determined thresholds. In case when thread’s queue is totally full a lock must be explicitly requested. Therefore, using batching updates approach we significantly lower the cost of lock contention, streamline access to locked structures and retain the precision of eviction algorithm such as LIRS. The key insight is that batching the updates on the eviction algorithm doesn’t materially affect the accuracy of the algorithm.

How are these ideas implemented in Infinispan? We introduced BoundedConcurrentHashMap class based on Doug Lea’s ConcurrentHashMap. BoundedConcurrentHashMap hashes entries based on their keys into lock protected segments. Instead of recording entries accessed per thread we record them in a lock free queue on a segment level. The main reason not to use ThreadLocal is that we could potentially have hundreds of threads hitting the data container, some of them very short lived thus possibly never reaching batching thresholds. When predetermined thresholds are reached eviction algorithms is executed on a segment level. Would running eviction algorithm on a segment level rather than entire data container impact overall eviction precision? In our performance tests we have not found any evidence of that.

Infinispan’s eviction algorithm is specified using strategy attribute of eviction XML element. In addition to old eviction approaches, starting with release 4.1.ALPHA2, you can now select LIRS eviction algorithm. LRU remains the default. Also note that starting with 4.1ALPHA2 release there are two distinct approaches to actually evict entries from the cache: piggyback and the default approach using a dedicated EvictionManager thread. Piggyback eviction thread policy, as it name implies, does eviction by piggybacking on user threads that are hitting the data container. Dedicated EvictionManager thread is unchanged from the previous release and it remains the default option. In order to support these two eviction thread policies a new eviction attribute threadPolicy has been added to eviction element of Infinispan configuration schema.

Does eviction redesign based on batching updates promise to live up to its expectations? Ding et al, authors of the original batching proposal, found that their framework increased throughput nearly twofold in comparison with unmodified eviction in postgreSQL 8.2.3. We do not have any numbers to share yet, however, initial testing of BoundedConcurrentHashMap were indeed promising. One of our partner companies replaced their crucial caching component with BoundedConcurrentHashMap and realized a 54% performance improvement on the Berlin SPARQL benchmark for their flagship product. Stay tuned for more updates.

Cheers,

Vladimir

Posted by Vladimir Blagojevic on 2010-03-30
Tags: eviction concurrency data structures

Tuesday, 12 May 2009

Implementing a performant, thread-safe ordered data container

To achieve efficient ordering of entries in the DataContainer interface for configurations that support eviction, there was a need for a linked HashMap implementation that was thread-safe and performant. Below, I specifically discuss the implementations of the FIFODataContainer and LRUDataContainer in Infinispan 4.0.x. Wherever this document references FIFODataContainer, this also applies to LRUDataContainer - which extends FIFODataContainer. The only difference is that LRUDataContainer updates links whenever an entry is visited as well as added.

After analysing and considering a few different approaches, the one I settled on is a subset of the algorithms described by H. Sundell and P. Tsigas in their 2008 paper titled Lock-Free Deques and Doubly Linked Lists, combined with the approach used by Sun’s JDK6 for reference marking in ConcurrentSkipListMap's implementation.

Reference marking? What’s that?

Compare-and-swap (CAS) is a common technique today for atomically updating a variable or a reference without the use of a mutual exclusion mechanism like a lock. But this only works when you modify a single memory location at a time, be it a reference or a primitive. Sometimes you need to atomically update two separate bits of information in a single go, such as a reference, as well as some information about that reference. Hence reference marking. In C, this is sometimes done by making use of the assumption that an entire word in memory is not needed to store a pointer to another memory location, and some bits of this word can be used to store additional flags via bitmasking. This allows for atomic updates of both the reference and this extra information using a single CAS operation.

This is possible in Java too using AtomicMarkableReference, but is usually considered overly complex, slow and space-inefficient. Instead, what we do is borrow a technique from ConcurrentSkipListMap and use an intermediate, delegating entry. While this adds a little more complexity in traversal (you need to be aware of the presence of these marker entries when traversing the linked list), this performs better than an AtomicMarkableReference.

In this specific implementation, the 'extra information' stored in a reference is the fact that the entry holding the reference is in the process of being deleted. It is a common problem with lock-free linked lists when you have concurrent insert and delete operations that the newly inserted entry gets deleted as well, since it attaches itself to the entry being concurrently deleted. When the entry to be removed marks its references, however, this makes other threads aware of the fact and cause CAS operations on the reference to fail and retry.

Performance

Aside from maintaining order of entries and being thread-safe, performance was one of the other goals. The target is to achieve constant-time performance - O(1) - for all operations on DataContainer.

The clhttp://3.bp.blogspot.com/_ca0W9t-Ryos/SgGwA2iw_vI/AAAAAAAAAKA/TpVMWo2Rq9U/s1600-h/FIFODataContainer.jpeg[image]ass diagram (click to view in full-size) depicts the FIFODataContainer class. At its heart the FIFODataContainer mimics a JDK ConcurrentHashMap (CHM), making use of hashtable-like lockable segments. Unlike the segments in CHM, however, these segments are simpler as they support a much smaller set of operations.

Retrieving data from the container The use of segments allow for constant-time thread-safe get() and containsKey() operations. Iterators obtained from the DataContainer - which implements Iterable, and hence usable in for-each loops - and keySet() are immutable, thread-safe and efficient, using traversal of the linked list - making use of getNext() and getPrev() helpers. See below for details. Traversal is efficient and constant-time.

Updating the container When removing an entry, remove() locks the segment in question, removes the entry, and unlinks the entry. Both operations are thread-safe and constant-time. Locking the segment and removing the entry is pretty straightforward. Unlinking involves marking references, and then an attempt at CAS’ing next and previous references to bypass the removed entry. Order here is important - updates to the next reference needs to happen first, read on for more details as to why.

When performing a put(), the entry is created, segment locked and entry inserted into the segment. The entry is then inserted at the tail of the linked list. Again, both operations are thread-safe and constant-time. Linking at the tail involves careful CAS’ing of references on the new entry, the tail dummy entry and the former last entry in the list.

Maintaining a lock-free, concurrent doubly linked list It is important to note that the entries in this implementation are doubly linked. This is critical since, unlike the JDK’s ConcurrentSkipListMap, we use a hashtable to look up entries to be removed, to achieve constant time performance in lookup. Locating the parent entry to update a reference needs to be constant-time as well, and hence the need for a previous reference. Doubly-linked lists make things much trickier though, as there two references to update atomically (yes, that sounds wrong!)

Crucially, what we do not care about - and do not support - is reverse-order traversal. This means that we only really care about maintaining accuracy in the forward direction of the linked list, and treat the previous reference as an approximation to an entry somewhere behind the current entry. Previous references can then be corrected - using the correctPrev() helper method described below - to locate the precise entry behind the current entry. By placing greater importance on the forward direction of the list, this allows us to reliably CAS the forward reference even if the previous reference CAS fails. It is hence critical that whenever any references are updated, the next reference is CAS’d first, and only on this success the previous reference CAS is attempted. The same order applies with marking references. Also, it is important that any process that touches an entry that observes that the next pointer is marked but the previous pointer is not, attempts to mark the previous pointer before attempting any further steps.

The specific functions we need to expose, to support DataContainer operations, are:

void linkAtEnd(LinkedEntry entry);

void unlink(LinkedEntry entry);

LinkedEntry correctPrev(LinkedEntry suggestedPrev, LinkedEntry current);

LinkedEntry getNext(LinkedEntry current);

LinkedEntry getPrev(LinkedEntry current);

These are exposed as protected final methods, usable by FIFODataContainer and its subclasses. The implementations themselves use a combination of CAS’s on a LinkedEntry’s next and previous pointers, marking references, and helper correction of previous pointers when using getNext() and getPrevious() to traverse the list. Note that it is important that only the last two methods are used when traversing rather than directly accessing a LinkedEntry’s references - since markers need to be stepped over and links corrected.

Please refer to Lock-Free Deques and Doubly Linked Lists for details of the algorithm steps.

Posted by Manik Surtani on 2009-05-12
Tags: algorithms eviction concurrency data structures

News

Tags

JUGs alpha as7 asymmetric clusters asynchronous beta c++ cdi chat clustering community conference configuration console data grids data-as-a-service database devoxx distributed executors docker event functional grouping and aggregation hotrod infinispan java 8 jboss cache jcache jclouds jcp jdg jpa judcon kubernetes listeners meetup minor release off-heap openshift performance presentations product protostream radargun radegast recruit release release 8.2 9.0 final release candidate remote query replication queue rest query security spring streams transactions vert.x workshop 8.1.0 API DSL Hibernate-Search Ickle Infinispan Query JP-QL JSON JUGs JavaOne LGPL License NoSQL Open Source Protobuf SCM administration affinity algorithms alpha amazon anchored keys annotations announcement archetype archetypes as5 as7 asl2 asynchronous atomic maps atomic objects availability aws beer benchmark benchmarks berkeleydb beta beta release blogger book breizh camp buddy replication bugfix c# c++ c3p0 cache benchmark framework cache store cache stores cachestore cassandra cdi cep certification cli cloud storage clustered cache configuration clustered counters clustered locks codemotion codename colocation command line interface community comparison compose concurrency conference conferences configuration console counter cpp-client cpu creative cross site replication csharp custom commands daas data container data entry data grids data structures data-as-a-service deadlock detection demo deployment dev-preview development devnation devoxx distributed executors distributed queries distribution docker documentation domain mode dotnet-client dzone refcard ec2 ehcache embedded query equivalence event eviction example externalizers failover faq final fine grained flags flink full-text functional future garbage collection geecon getAll gigaspaces git github gke google graalvm greach conf gsoc hackergarten hadoop hbase health hibernate hibernate ogm hibernate search hot rod hotrod hql http/2 ide index indexing india infinispan infinispan 8 infoq internationalization interoperability interview introduction iteration javascript jboss as 5 jboss asylum jboss cache jbossworld jbug jcache jclouds jcp jdbc jdg jgroups jopr jpa js-client jsr 107 jsr 347 jta judcon kafka kubernetes lambda language learning leveldb license listeners loader local mode lock striping locking logging lucene mac management map reduce marshalling maven memcached memory migration minikube minishift minor release modules mongodb monitoring multi-tenancy nashorn native near caching netty node.js nodejs nosqlunit off-heap openshift operator oracle osgi overhead paas paid support partition handling partitioning performance persistence podcast presentations protostream public speaking push api putAll python quarkus query quick start radargun radegast react reactive red hat redis rehashing releaase release release candidate remote remote events remote query replication rest rest query roadmap rocksdb ruby s3 scattered cache scripting second level cache provider security segmented server shell site snowcamp spark split brain spring spring boot spring-session stable standards state transfer statistics storage store store by reference store by value streams substratevm synchronization syntax highlighting testing tomcat transactions tutorial uneven load user groups user guide vagrant versioning vert.x video videos virtual nodes vote voxxed voxxed days milano wallpaper websocket websockets wildfly workshop xsd xsite yarn zulip

back to top