bwyogatama / mordred

Mordred

Mordred is a research prototype of a heterogeneous CPU-GPU DBMS engine. The work was presented at VLDB '22. Please read the paper for more details.

@article{10.14778/3551793.3551809,
author = {Yogatama, Bobbi W. and Gong, Weiwei and Yu, Xiangyao},
title = {Orchestrating Data Placement and Query Execution in Heterogeneous CPU-GPU DBMS},
year = {2022},
issue_date = {July 2022},
publisher = {VLDB Endowment},
volume = {15},
number = {11},
issn = {2150-8097},
url = {https://doi.org/10.14778/3551793.3551809},
doi = {10.14778/3551793.3551809},
journal = {Proc. VLDB Endow.},
month = {jul},
pages = {2491–2503},
numpages = {13}
}

Mordred is currently under development. Bug fixes will be released soon!!!

Reach out to [email protected] for questions or issues.

Usage

*Dependencies:

Ubuntu > 20.04
GCC > 9.4
CUDA > 11.0
IntelTBB

To use Mordred:

To run the Star Schema Benchmark implementation:

• Generate the test dataset

cd test/

# Generate the test generator / transformer binaries
cd ssb/dbgen
make
cd ../loader
make 
cd ../../

# Generate the test data and transform into columnar layout
# Substitute <SF> with appropriate scale factor (eg: 1)
python util.py ssb <SF> gen
python util.py ssb <SF> transform

#Sort the LINEORDER table if you want to enable segment skipping
cd ssb/loader
make sort
./columnSort ../data/s{SF}_columnar/LINEORDER ../data/s{SF}_columnar/LINEORDERSORT 5 16 {COLUMN_SIZE}

• Configure the benchmark settings

# Edit SF and BASE_PATH in src/gpudb/ssb_utils.h
# Edit SF and BASE_PATH in src/gpudb/common.h
# Edit file path in src/gpudb/minmax.cpp and src/gpudb/minmaxsort.cpp
# Edit SF and table size in minmax.sh (following the size in src/gpudb/ssb_utils.h)

• To compile and run Mordred

make setup
make bin/gpudb/minmax
make bin/gpudb/minmaxsort
./minmax.sh
make bin/gpudb/main
./bin/gpudb/main

When you run Mordred, you would be given the following options:

Select Options:
1. Run Specific Query
2. Run Random Queries
3. Run Experiment
4. Replacement
5. Dump Trace
6. Exit
cache. Cache Specific Column
clear. Delete Columns from GPU
custom. Toggle custom malloc
skipping. Toggle segment skipping

Mordred do not cache the data by default To populate the cache, select option (2), run 100 random queries (for warm-up), and then select option (4) to populate the cache.

In option (4), you can input the caching policies from the following options: (1)LFU, (2)LRU, (3)LRU2, (4)LFUSegmented, (5)LRUSegmented, (6)LRU2Segmented, and (7)SemanticAware. SemanticAware is the novel caching policy introduced in our paper. See Figure 7 in our paper to understand the difference between each policies.

To run individual query, select option (1) and input the ssb query (e.g. input '21' to run q2.1 in SSB).

mordred's Issues

Warp out-of-range error while running cache replacement experiment.

Hi,

I am seeing the following error while running an experiment where we can choose the cache replacement policy. I will discuss the config case with the cache replacement policy set to LRU and the query index is 2.2. Please note, this error is persistent across other replacement policies and queries.

Where is the error:
File name: src/gpudb/BlockLibrary.cuh
Function: BlockBuildValueGPUDirect()
Line: atomicCAS(&ht[hash << 1], 0, vals[ITEM]);

What is the error:
CUDA_EXCEPTION_5, Warp Out-of-range Address.

Reason for getting the error:
hash << 1 is out of bounds for ht. In the experiment mentioned above, I see that the hash generated is 4294967295, while ht_len is 200000. Due to this, there is a warp out of bound error.

Can you please help me debug this issue?

Thanks in advance.