Efficient metadata caching

In my last post I was talking about 1 Million mailboxes. Each of them is a directory with several subdirectories, like Trash, Sent Items etc.

The mailbox directory itself lies 4 directories below the root node (like /a/b/c/mailbox). The hierarchy is managed by our mail-store application.

I don't know the average number of files/directory, but let's assume, each mailbox consists in average of 20 files/directories, we would currently have about 20'000'000 inodes.

Mailbox access is mostly random. We don't know when a mail is coming in, we also don't know about when a user is reading his mails. What we now from experience is, that a lot of time is spend in looking up metadata.

With mostly random access (we measured it as ~ 55% write / 45% read a while ago), and the amount of data, the chance to identify a data working set is quite low. Ok, maybe recently received emails could be part of a "working-set".

But wouldn't it be great if we could cache as much metadata as possible?

Roch Bourbonnais wrote a blog entry a while ago about inodes on zfs. This is by no means a scientific analysis, but let's take his numbers:

"23.8M files consuming 27GB of data. Basically less than 1.2K of used disk space per KB of files"

Let's say, each mail/directory uses 0.2K, and we have 20'000'000 of them, we would currently have 3.8GB of inode data. No problem to cache that. I certainly have to investigate a little bit more what kind of metadata the ARC Cache is caching.

Thanks to analytics, I can at least do a bit of sanity checking, it currently shows me that around 11G inside ARC are used for metadata caching.



If we take another view, we can see that not only do we have metadata cached, it is also heavily used. In this picture I have colored all cache hits.



Lessons learned today:

-ZFS does not waste space for inodes and therefore not cache.
-ARC is very efficient

Questions to be answered:

-What does "metadata" include?

Comments

Post a Comment

Popular posts from this blog

SLOG Latency

Image
After reading Brendan's newest blog entry, I was curious about what kind of slog latency we can see for our data migration load. To remind you, only synchronous writes go into the slog SSD devices. As this is a migration running, we can see mostly NFS write operations: In our configuration the SSD slog is mirrored (HDD4 and HDD8). Hence the same number of IOPS: The next picture shows us the latency for our SDD SLOG Device HDD 4. We can see here that latencies start at 79 us and are mostly under 200 us. There are some outliers, but approx. 95% are under 500 us: This matches quite well with the values Brendan blogged about (137-181 us), which includes NFSv3 latency. For reference (no picture here), we can see latencies of about 170-500 us mostly for NFSv4. By the way. SLOG Devices are mostly one way devices, as shown here. Only if things go really bad, they are read from...
Read more

Heating up the Data Pipeline (Part 1)

Image
Pre-Processing Data I often hear the question from our customers, how data can be transformed prior to indexing in Splunk. Damien from  Baboonbones  has done a tremendous job in creating add-ons   providing custom inputs for Splunk. Most of his custom inputs provide the means to pre-process data by allowing custom event handlers to be written. Sometimes you still want to pre-process data that gets collected from Splunk's standard input types, like file monitors, Windows EventLogs, scripted inputs etc. Also, not everyone is capable of writing custom event handlers. A requirement these customers have, is that they have rolled out a large number of Splunk Universal Forwarders and they do not want to install another agent. To summarize, the solution capable of pre-processing data, should be easy to use, be easily integrated and be build on top of their existing architecture. How to plumb Splunk Pipelines Splunk has its own fittings to connect a Universal Forwar
Read more

Heating up the Data Pipeline (Part 2)

Image
In Part 1 we went through how to route events from Splunk to a 3rd party system without losing metadata. Now I'll show you how events can be transformed using Apache NiFi and be sent back to Splunk into the  HTTP Event Collector . Note: The following is not a step-by-step documentation. To learn how to use Apache NiFi you should read the G etting Started Guide . Simple Pass-Through Flow As a first exercise we will create a simple flow, that only passes data through NiFi, without applying any complex transformations. The following picture shows a high-level NiFi flow, that receives events in our custom uncooked event format with a TCP listener, then sends the data further into a transformation "black box" (aka Processor Group), which emits events in a format, that can be ingested into a Splunk HTTP Event Collector input. Apache Nifi currently provides a rapidly growing number of processors (currently 266), which can be used for data ingestion, transfo
Read more