Scalability, in Graphical Form, Analyzed

I'm at Surge, this week, where I just listened to Baron Schwartz give a talk about scalability and performance.  As usual, Baron was careful to distinguish between performance (which is how fast it is) and scalability (which is how much faster you can make it by adding more resources).  One of the things Baron talked about was Neil Guenther's Universal Scalability Law, which attempts to model the behavior of complex systems (such as database systems) as you add threads, or nodes, or users.

Postgres Open, and What Makes a Good Talk

My last few blog posts have been all about scalability and performance, mostly because that's what I've been spending most of my time on these last few months.  But I'm pleased to have another subject: Postgres Open.  I've been going to PostgreSQL conferences for a few years now, but Postgres Open - which is a new conference this year - was my first experience being on the program committee.

Index-Only Scans: Now There's a Patch

In November of 2010, I blogged about a much-requested PostgreSQL feature: index-only scans.  We've made some progress!  In June of this year, I committed a patch (and then, after Heikki found some bugs, another patch) to make the visibility map crash-safe.  In previous releases, it was possible for the visibility map to become incorrect after a system crash, which means that it could not be relied on for anything very critical.  That should be fixed now.  Last week, I posted a patch for the main feature: index-only scans.

Linux and glibc Scalability

As some of you probably already know from following the traffic on pgsql-hackers, I've been continuing to beat away at the scalability issues around PostgreSQL.  Interestingly, the last two problems I found turned out, somewhat unexpectedly, not to be internal bottlenecks in PostgreSQL.  Instead, they were bottlenecks with other software with which PostgreSQL was interacting during the test runs.

More On Read Scaling

In my previous blog post on read scaling out to 32 cores, I wrote a patch I recently committed to improve read scalability in PostgreSQL.  It turns out, however, that there's a downside to that patch, which was uncovered in testing by Stefan Kaltenbrunner.  (Fortunately, I have a fix.)  And, there's an opportunity for further performance improvement by applying a similar technique to an additional type of lock.  For full details, read on.

Read Scaling Out to 32 Cores

With the exception of a week's vacation, my last month has been mostly absorbed by PostgreSQL performance work.  Specifically, I've been looking at the workload generated by "pgbench -S", which essentially fires off lots and lots of SELECT queries that all do primary key lookups against a single table.  Even more specifically, I've been looking at the way this workload performs on systems with many CPU cores where (I found) PostgreSQL was not able to use all of the available CPU time to answer queries. Although the single-core performance was around 4,300 tps, performance with 36 clients (on a 24-core server) was only about 36,000 tps.

Research revealed that performance was being limited mostly by PostgreSQL's lock manager.  Each SELECT query needed to lock the table being queried - and its index - against a concurrent DROP operation.   Since PostgreSQL 8.2, the lock manager has been partitioned: a lock request against a database object will be assigned to one of 16 "partitions", and lock requests against objects that fall into different partitions can proceed in parallel.  Unfortunately, that doesn't help much in this case, because only two objects are being locked: the table, and its index.  Most of the traffic therefore targets just two of the sixteen lock manager partitions.  Furthermore, because the query itself is so trivial, the rate of lock and unlock requests is extremely high - on a more complex query, the bottleneck wouldn't be as severe.

False Contention

For the last few weeks, I've been buried in PostgreSQL performance analysis, mostly focusing on the workload generated by "pgbench -S" under high concurrency.  In other words, lots and lots of very simple SELECT statements on a single table.  Such workloads can generate serious internal contention within PostgreSQL on systems within many CPU cores.

But it's interesting to note that most of the contention points I've so far identified are what might be called "false contention".  The transactions generated by this workload need to perform operations such as:

- acquire an AccessShareLock on the target relation and its index to guard against a concurrent drop or schema change
- acquire an ExclusiveLock on their VXID in case another transaction wishes to wait for transaction end
- read the list of pending "invalidation" events, which are generally created by DDL
- read the list of in-progress XIDs, to generate a snapshot
- find the root index block for the table's primary key index in the shared buffer pool
- momentarily pin the block containing the root index page into the buffer pool, so that it can't be evicted while we're examining it
- momentarily lock the root index page, so that no new items can be added until we've decided which downlink to follow

Now, in fact, in this workload, there is no concurrent DDL against the relevant table and index, or any other one; no one is attempting to wait for the completion of any VXID; the list of in-progress XIDs can easily be simultaneously read by many backends; and the root index page is in no danger either of being modified or of being evicted.  In short, the problem is not that there are resource conflicts, but that verifying that no resource conflicts exist is itself eating up too many resources.