Plasma_release_0_3

Release Notes For Plasma

This is version: 0.3 "Einmaleins". This is an alpha release to make Plasma known to interested developers.

Changes

Changed in 0.3

New features in PlasmaFS:

• Name nodes can now also be referenced by IP addresses
• Added Plasma_netfs module for easy access to PlasmaFS files
• Improved plasma command-line utility (wildcards, tree mounting)

Implementation improvements in PlasmaFS:

• The representation of blocklists has been optimized. Whereever possible, a compressed format is used now, including the database table. This is assumed to fix performance problems for files with long blocklists.
• Improved block allocation algorithm: it is tried to avoid discontiguities as much as possible
• The DB accesses are now completely non-blocking
• Access tickets are allocated in advance by the namenode
• Enhanced versions of Plasma_client.copy_in and copy_out allow buffered reading and writing to some extent
• Better scheme for emitting debug log messages

New features in the map/reduce framework:

• It is now possible to execute several jobs at the same time. However, the jobs must use the same task server, and hence the same m/r algorithm. (The latter is no problem for streaming jobs.)

Implementation improvements in the map/reduce framework:

• Optimized record reader
• Optimized sorting algorithm
• If RAM is tight, the execution of tasks can be delayed

Improvements for both PlasmaFS and map/reduce:

• RAM management: The RAM for large buffers and shared memory buffers is actively managed, and this works now in a fairly automatic way

Compatibility:

• Existing PlasmaFS filesystems are incompatible (db schema changes)
• There are incompatible protocol changes

What is working and not working in PlasmaFS

Generally, PlasmaFS works as described in the documentation. Crashes have not been observed for quite some time now, but occasionally one might see critical exceptions in the log file.

PlasmaFS has so far only been tested on 64 bit, and only on Linux as operation system. There are known issues for 32 bit machines, especially the blocksize must not be larger than 4M.

Data safety: Cannot be guaranteed. It is not suggested to put valuable data into PlasmaFS.

Known problems:

• It is still unclear whether the timeout settings are acceptable.
• There might be name clashes for generated file names. Right now it is assumed that the random number generator returns unique names, but this is for sure not the case.
• The generated inode numbers are not necessarily unique after namenode restarts.
• The cp subcommand of the plasma utility is slower than necessary. Avoid it for now! Alternatives: Use the get, put, and cat subcommands instead.

Not implemented features:

• The namenodes cannot yet detect crashed datanodes. Datanodes are always reported as alive.
• The ticket system is not fully implemented (support for "read").
• There is no authorization system (file access rights are ignored)
• There is no authentication system to secure filesystem accesses (such as Kerberos)
• There are too many hard-coded constants.
• The file name read/lookup functions should never return ECONFLICT errors. (This has been improved in 0.2, though.)
• Translation of access rights to NFS
• Support for checksums
• Support for "host groups", so that it is easier to control which machines may store which blocks. Semantics have to be specified yet.
• Define how blocks are handled that are allocated but never written.
• Recognition of the death of the coordinator, and restart of the election algorithm.
• Multicast discovery of datanodes
• Lock manager (avoid that clients have to busy wait on locks)
• Restoration of missing replicas
• Rebalancing of the cluster
• Automated copying of the namenode database to freshly added namenode slaves

What is working and not working in Plasma MapReduce

Not implemented features:

• Task servers should be able to provide several kinds of jobs
• Think about dynamically extensible task servers
• Run jobs only defining map but no reduce.
• Support for combining (an additional fold function run after each shuffle task to reduce the amount of data)
• nice web interface
• support user counters as in Hadoop
• restart/relocation of failed tasks
• recompute intermediate files that are no longer accessible due to node failures
• Speculative execution of tasks
• Support job management (remember which jobs have been run etc.)

What we will never implement:

• Jobs only consisting of reduce but no map cannot be supported due to the task scheme. (Reason: Input files for sort tasks must not exceed sort_limit.)

Changed in 0.3.1

Release 0.3.1 was tested in a configuration with four nodes and lots of RAM. The encountered difficulty was that the periodic fsync call took that long that client requests timed out. The following measures improved that:

• Increasing general timeouts in Plasma client
• Changed flow control in datanode servers: Read requests are never throttled - it is assumed the OS throttles automatically. Write requests are still throttled if the fsync syscall takes long.
• Datanode servers: Local write requests are now included in the throttling implementation.
• Added a statistics report to the datanode servers

Other bugfixes:

• Wrong parsing of the alive_min config parameter
• Tickets: Avoiding a race between a reset request and the normal housekeeping loop

Also:

• Increasing the backlog for all servers in the standard config

This web site is published by Informatikbüro Gerd Stolpmann