Working with large datasets¶
You’ll have noticed in setting up Camelot, that you probably did not need to provision servers, configure relational databases, network equipment and so on. Camelot does not treat working with massive volumes of data as its priority use case. However, we’ve spent much time ensuring you won’t encounter growth pain, no matter how much data you want to throw at it (within reason, of course).
This page describes what you can expect, and recommendations should you be dealing with large datasets.
Camelot has no known upper limit on the amount of data it can support, however some parts of Camelot will take longer to load as the size of the dataset grows. The authors have simulated datasets with 2 million images to ensure that Camelot will perform well for 99% of datasets Camelot could be used for.
The main performance considerations with this volume of data is performing the initial load of the library, searching the library, and the CPU and memory constraints required to produce reports.
To serve as a guide for how Camelot may perform for large datasets, below is the timing for working with a dataset of 2 million images using a high-end laptop in 2017 (Dell Precision 5510; 16GB RAM; Intel i7-6820HQ; SSD) with a maximum JVM heap size of 14GB:
- Full Export report: 368 seconds
- Library load time: 13 seconds
- Library search time (basic search): 6.5 seconds
- Library search time (full-text search): 28 seconds
Considerations for your dataset¶
This section will describe considerations relevant for using Camelot with larger datasets on a Camelot server, and offer some guidelines for what the authors would expect to be reasonable configurations under various scenarios.
Data for Camelot’s performance is known for only a small number of scenarios, thus the numbers offered are very much guidelines and you may find different considerations important in your use case.
Memory can be a difficult consideration to judge. Generally speaking, while there is sufficient memory, it has little impact upon performance. However when there is not enough memory, it may result in things which will take considerably longer to complete, or may never complete. There are two main aspects to memory in Camelot: the physical memory available, and the memory available to the JVM heap.
The most important consideration is maximum size of the JVM heap. Regardless of how much physical memory is available to a machine, Camelot may only use the memory available in the JVM heap (and some small additional amounts, as per other JVM configurations), and this is something should be configured manually for large datasets.
What you can expect without configuration¶
By default, Java will (typically) make available 25% of the total memory available in your machine to Camelot. This gives the following approximate minimum memory requirements for various report sizes:
|Dataset size (images)||Heap size (MB)||Physical memory (MB)|
From Camelot’s perspective, the important column here is the Heap Size, which is the amount of memory which it can actually use. However, Camelot can use any and all available physical memory with the appropriate configuration of Java.
Configuring the Heap Size becomes increasingly important for resource-efficient use of Camelot.
The JVM heap size should not exceed the size of physical memory available, and ideally should not impinge upon the resources required by other applications on the machine Camelot is running upon.
Use the administration interface to set the heap size to your needs.
As a (very) rough guide, the authors suggest an additional 800MB of heap space for every 100,000 images, with a starting heap space of 1GB.
High throughput and low latency storage for a database is always nice to have, though should not be strictly necessary to use Camelot with large datasets. However, using Camelot with remote storage, for example, running Camelot on a laptop and connecting to a database stored on a NAS over wifi, is unlikely to result in nice performance characteristics. Effort should be made to reduce the latency and maximise throughput between where Camelot is running, and where its database resides.
That said, much of the volume of data for an installation of Camelot is occupied by images. Images may reside on lower-throughput and higher latency storage as it does not impact significantly on Camelot’s performance profile.
We currently do not have sufficient data on how the number of concurrent users impacts upon Camelot’s performance in Real World usage. It depends heavily on the usage of those users. Should you be using Camelot with a large number of users, and have any questions or feedback about your particular use case, please do reach out via the forum.
This section applies to client machines: those connecting to a Camelot server, which do not run a copy of Camelot themselves.
Generally speaking, any computer able to achieve an acceptable degree of responsiveness should be a fine candidate for accessing Camelot running on a remote machine. The main consideration of client machines is less-so performance, than it is screen resolution. In common usage, there should be no discernible degradation on performance for large datasets for client machines.