Sorry I’m late to this discussion. I think that the motivation is correct. There is really quite a bit of activity around this issue. Let’s take extra efforts to engage extra time with commits to confirm performance improvements.
Let’s particularly pay attention to threading. +1 Regards, Dave Sent from my iPhone > On Jul 21, 2022, at 11:37 AM, Matteo Merli <mme...@apache.org> wrote: > > ## Motivation > > The current implementation of the read cache in the Pulsar broker has largely > remained unchanged for a long time, except for a few minor tweaks. > > While the implementation is stable and reasonably efficient for > typical workloads, > the overhead required for managing the cache evictions in a broker > that is running > many topics can be pretty high in terms of extra CPU utilization and on the > JVM > garbage collection to track an increased number of medium-lived objects. > > The goal is to provide an alternative implementation that can adapt better to > a wider variety of operating conditions. > > ### Current implementation details > > The broker cache is implemented as part of the `ManagedLedger` component, > which sits in the Pulsar broker and provides a higher level of > abstraction of top > of BookKeeper. > > Each topic (and managed-ledger) has its own private cache space. This > cache is implemented > as a `ConcurrentSkipList` sorted map that maps `(ledgerId, entryId) -> > payload`. The payload > is a `ByteBuf` reference that can either be a slice of a `ByteBuf` that we got > when reading from a socket, or it can be a copied buffer. > > Each topic cache is allowed to use the full broker max cache size before an > eviction is triggered. The total cache size is effectively a resource > shared across all > the topics, where a topic can use a more prominent portion of it if it > "asks for more". > > When the eviction happens, we need to do an expensive ranking of all > the caches in the broker > and do an eviction in a proportional way to the currently used space > for each of them. > > The bigger problem is represented by the `ConcurrentSkipList` and the > `ByteBuf` objects > that need to be tracked. The skip list is essentially like a "tree" > structure and needs to > maintain Java objects for each entry in the cache. We also need to > potentially have > a huge number of ByteBuf objects. > > A cache workload is typically the worst-case scenario for each garbage > collector implementation because it involves creating objects, storing > them for some amount of > time and then throwing them away. During that time, the GC would have > already tenured these > objects and copy them into an "old generation" space, and sometime > later, a costly compaction > of that memory would have to be performed. > > To mitigate the effect of the cache workload on the GC, we're being > very aggressive in > purging the cache by triggering time-based eviction. By putting a max > TTL on the elements in > the cache, we can avoid keeping the objects around for too long to be > a problem for the GC. > > The reverse side of this is that we're artificially reducing the cache > capacity to a very > short time frame, reducing the cache usefulness. > > The other problem is the CPU cost involved in doing these frequent > evictions, which can > be very high when there are 10s of thousands of topics in a broker. > > > ## Proposed changes > > Instead of dealing with individual caches for each topic, let's adopt > a model where > there is a single cache space for the broker. > > This cache is broken into N segments which act as a circular buffer. > Whenever a segment > is full, we start writing into the next one, and when we reach the > last one, we will > restart recycling the first segment. > > This model has been working very well for the BookKeeper `ReadCache`: > https://github.com/apache/bookkeeper/blob/master/bookkeeper-server/src/main/java/org/apache/bookkeeper/bookie/storage/ldb/ReadCache.java > > The eviction becomes a completely trivial operation, buffers are just > rotated and > overwritten. We don't need to do any per-topic task or keep track of > utilization. > > Today, there are 2 ways of configuring the cache, one that "copies" > data into the cache > and another that will just use reference-counting on the original > buffers to avoid > payload copies. > > ### Memory copies into the cache > > Each segment is composed of a buffer, an offset, and a hashmap which maps > `(ledgerId, entryId) -> offset`. > > > The advantage of this approach is that entries are copied into the cache > buffer > (in direct memory), and we don't need to keep any long-lived Java objects > around > > ### Keeping reference-counted buffers in the cache > > Each segment in the cache will contain a map `(ledgerId, entryId) -> ByteBuf`. > Buffers will have an increase reference count that will keep the data > alive as long > as the buffer is in the cache and it will be released when the cache > segment is rotated. > > The advantage is we avoid any memory copy when inserting into or > reading from the cache. > The disadvantage is that we will have references to all the `ByteBuf` > objects that are in the cache. > > ### API changes > > No user-facing API changes are required. > > ### New configuration options > > The existing cache implementation will not be removed at this point. Users > will > be able to configure the old implementation in `broker.conf`. > > This option will be helpful in case of performance regressions would be seen > for > some use cases with the new cache implementation. > > > > -- > Matteo Merli > <mme...@apache.org>
