create new tag
view all tags

Meeting Minutes - 2011-05-17

Attendants: Wenjing, Ran, Jie, Fabio

Time: 2:30 p.m.

Secretary: Fabio

  • OpenStack Swift [Ran]
    • Swift deployed in the testbed over 6 virtual machines. However, the limited free disk space on those machines (currently 4 GB) does not allow for realistic tests.
    • The current deployment allowed Ran to get familiar with the components of the product and the provided tools for administering and monitoring the system and also for uploading, downloading files (up to 5GB per file) and for creating and deleting directories.
    • There is also an API provided for developing applications backed by Swift.
    • Next steps are to perform some meta-data centered tests (i.e. by creating small files) and also test the provided Python API.
  • Cassandra [Wenjing]
    • Wenjing prepared a presentation summarizing her tests using Cassandra as a meta-data store for a file hierarchy. Her presentation is stored here.
    • for this test, one Cassandra keyspace was created containing 2 column families: one for storing information about the directories and one for files. Each object (file or directory) is uniquely identified within the store by a UUID. Each entry in the the column family for Directories contains the information associated to one directory (which key is a UUID) and has columns for storing the owner and group of the directory and one column per object (file or directory) contained in it.
    • the information for one file is stored in the File column family and contains columns for storing the file meta-data (create/modify/access time, owner, group, file size) and file contents.
    • using the Python client (which is based on the THRIFT framework provided by Cassandra), Wenjing developed for this test a set of tools for executing operations on the file store (such as list files & directories, make directory, rename a file or directory, upload and download files).
    • given this representation, navigating the hierarchy is straightforward: one example is given in the presentation (see slides 21 and 22).
    • an initial test were performed using a Cassandra cluster of 4 virtual machines on the test platform, using a replication factor of 3. The test was focused on metadata operations and basically consisted of commands equivalent to 'mkdir', 'touch file', 'ls' with a hierarchy depth of 4 levels. There were more than 250.000 operations and each one took on average 0.4 seconds. This performance seems a priori low and needs to be compared to an equivalent test on other backend storage systems (Lustre, GPFS, dCache, etc.). It is important to bear in mind that this was the first test and there were no tuning at all nor in the Cassandra cluster nor on the virtual machines used in the testbed platform.
    • another test which tried to perform 10 million operations crashed due to a memory problem not yet understood.
  • Discussion [all but Ran who had to left earlier]
    • it appears that the current setup of the testbed platform is not adequate for the kind of tests we want to perform. Fabio will find some ways to have more appropriate hardware. In the meantime, we decided to reduce the number of virtual machines per host machines to 3 (instead of 4 as currently) to make those virtual machines more usable.
    • we need also to install the Ganglia probes for collecting data on the virtual machines behavior during the tests and understand what is really going on with the system.
    • we also discussed the question of the suitability of the key-value stores for the purpose of building a FileStore, as opposed to use a more conventional distributed file system (Lustre, Gluster, etc.) or a high-level software layer (such as dCache). No definitive answers found. It is believed that the key-value stores are very promising and more work on the FileStore project is needed to answer that question.
  • Next meeting: Wednesday June 15th 2011, 10h30, Fabio's office.

-- FabioHernandez - 2011-05-18

Topic revision: r1 - 2011-05-18 - FabioHernandez
This site is powered by the TWiki collaboration platform Powered by PerlCopyright © 2008-2022 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding TWiki? Send feedback