bnclabs / llrb-index Goto Github PK

Performance measurement for Llrb instances. Parameters to measure:

• Latency of a operations in min, max, average and in percentile.
• Throughput of operations.

These measurements are to taken with 3 sets of data -
1M data set, 10M data set and 100M data set.

This performance suite comprises of:

• Initial load, 1M and 10M and 100M
• Set operation, overwrite 1M data set, overwrite 10M and overwrite 100M data set
• Delete operation, reduce 10M data set to 1M data set and reduce 100M data set to 10M data set
• Get operation, get all keys in 1M, 10M, 100M data sets.

Measuring:

• Latency, for every 10th operation measure the latency.
• Throughput, compute the total time taken to complete all the operation.
• Tally, latency * num_op = throughtput

Repeat the entire suite for:

• {key=i64, value=i64}
• {key=String(32), value=String(256)}

Precreate the set before applying the operation.

Right now, iter(), range() and reverse() APIs return iterators for full-table scan, forward and reverse iteration between a low and high bound.

These APIs are kept separate for performance reasons. It is also possible to unify these APIs into a DoubleEndedIterator for user convenience.

In case you need llrb-index to be part of a proprietary distribution or to be part of a different license agreement, please post your request here, or contact the author directly.

Create is an essential part of CRUD access to data indexes. Something like:

fn create(&mut self, key: K, value: V);

Depth calculates minimum, maximum, average and percentile of leaf-node's depth in the Llrb tree. Write unit test cases under depth_test.rs for Depth type.

When types like Vec< T> and String are used as key, it may not be cache friendly. Figure out a way to retain parametrised Llrb type and co-locating atleast Vec< u8> with Node< K,V>