IOQP

An Impact Ordered Query Processor.

Credits

IOQP is loosely based on the JASS and JASSv2 search systems.

IOQP makes use of open-source Rust code from the faster-graph-bisection library.

Citation Information

If you use this code in your own work or research, please consider citing our work:

@inproceedings{mpg22-desires,
 title = {IOQP: A simple Impact-Ordered Query Processor written in Rust},
 author = {J. Mackenzie and M. Petri and L. Gallagher},
 booktitle = {Proc. DESIRES},
 year = {2022},
 pages = {22--34},
}

Build instructions

You can build the code using cargo:

cargo build --release

Running Experiments

Use the following scripts to run the Gov2 and MS MARCO experiments from the paper:

./script/download-data.sh
./script/build-indexes.sh
./script/run-queries.sh

The run files are located in data/gov2/runs and data/msmarco/runs. Timing results can be found in data/log.

Throughput Experiments

Here we give an example of the multi-threaded throughput experiments for Gov2. In the following we first start the server to host an index and listen for incoming requests. Then we run the load generator to simulate a workload of incoming queries.

1. Starting the server

$ ./target/release/serve --max-blocking-threads 16 --index data/gov2/indexes/bp-gov2.8.ioqp.idx 
2022-08-17T01:19:18.406794Z  INFO serve: args = Args { index: "data/gov2/indexes/bp-gov2.8.ioqp.idx", port: 3000, max_blocking_threads: 16 }
2022-08-17T01:19:18.406844Z  INFO serve: loading index from file data/gov2/indexes/bp-gov2.8.ioqp.idx
2022-08-17T01:21:22.277424Z  INFO serve: start http endpoint at 0.0.0.0:3000

2. Run the load generator. In this example we are using exhaustive processing with 10 incoming queries per second.

$ ./target/release/load_gen --k 1000 --mode fraction-1 --queries data/gov2/queries/mqt.queries --tps 10
2022-08-17T01:29:59.276151Z  INFO load_gen: read queries = 59986
2022-08-17T01:34:59.322531Z  INFO load_gen: ======= Server Time =======
2022-08-17T01:34:59.322544Z  INFO load_gen: # of samples: 2009
2022-08-17T01:34:59.322548Z  INFO load_gen:   50'th percntl.: 30500µs
2022-08-17T01:34:59.322553Z  INFO load_gen:   90'th percntl.: 69395µs
2022-08-17T01:34:59.322558Z  INFO load_gen:   99'th percntl.: 116047µs
2022-08-17T01:34:59.322562Z  INFO load_gen: 99.9'th percntl.: 149946µs
2022-08-17T01:34:59.322565Z  INFO load_gen:             max.: 164049µs
2022-08-17T01:34:59.322578Z  INFO load_gen:        mean time: 36120.8µs
2022-08-17T01:34:59.322648Z  INFO load_gen: ======= User Time =======
2022-08-17T01:34:59.322651Z  INFO load_gen: # of samples: 2009
2022-08-17T01:34:59.322655Z  INFO load_gen:   50'th percntl.: 31352µs
2022-08-17T01:34:59.322660Z  INFO load_gen:   90'th percntl.: 70331µs
2022-08-17T01:34:59.322663Z  INFO load_gen:   99'th percntl.: 116714µs
2022-08-17T01:34:59.322666Z  INFO load_gen: 99.9'th percntl.: 150643µs
2022-08-17T01:34:59.322669Z  INFO load_gen:             max.: 164848µs
2022-08-17T01:34:59.322673Z  INFO load_gen:        mean time: 36953.0µs

Data

You will require a unique URL with a password to get the data due to the old data storage platform being decommissioned. Please create a github issue or contact Joel directly: [email protected]

Indexing

Use the script/build-indexes.sh to build the indexes from the paper.

Index a CIFF file and perform quantization:

./target/release/create \
    --input data/gov2/ciff/bp-gov2.ciff \
    --output data/gov2/indexes/bp-gov2.8.ioqp.idx \
    --quantize \
    --quant-bits 8 \
    --bm25-k1 0.9 \
    --bm25-b 0.4

Index a CIFF file that is already quantized:

./target/release/create \
    --input data/msmarco/ciff/bp-spladev2.ciff \
    --output data/msmarco/indexes/bp-spladev2.ioqp.idx

Query processing

Use the script/run-queries.sh to run the queries from the paper.

Query processing with exhaustive mode:

./target/release/query \
    --index data/gov2/bp-gov2.8.ioqp.idx \
    --queries data/gov2/queries/gov2.queries \
    --output data/gov2/run/gov2.run \
    --k 1000 \
    --mode fraction-1 \
    --warmup

Query processing with fixed budget:

./target/release/query \
    --index data/gov2/bp-gov2.8.ioqp.idx \
    --queries data/gov2/queries/gov2.queries \
    --output data/gov2/run/gov2.run \
    --k 1000 \
    --mode fixed-10000 \
    --warmup

Query processing with query term weights:

./target/release/query \
    --index data/msmarco/bp-spladev2.ioqp.idx \
    --queries data/msmarco/queries/spladev2.dev.query \
    --output data/msmarco/run/spladev2.run \
    --k 1000 \
    --weighted