Riveter

An adaptive query suspension and resumption framework. It supports redo, pipeline-level, process-level query suspension and resumption strategy.

Datasets

We use TPC-H and TPC-DS benchmark to generate datasets.

TPC-H Dataset

First, generating the original tables (tbl format) using TPC-H tools, and simply running duckdb_tpch_data.py can convert the table files to parquet or csv format using the following command,

# make sure in the tpch folder
cd tpch
python3 duckdb_tpch_data.py -d ../dataset/tpch/tbl-sf1 -f parquet -rgs 10000

You can move the converted data to any folder you want.

We have several datasets:

• TPC-H SF-0.01 (Tiny): dataset/tpch/tbl-tiny, dataset/tpch/parquet-tiny
• TPC-H SF-0.1 (Small): dataset/tpch/tbl-small, dataset/tpch/parquet-small
• TPC-H SF1: dataset/tpch/tbl-sf1, dataset/tpch/parquet-sf1
• TPC-H SF10: dataset/tpch/tbl-sf10, dataset/tpch/parquet-sf10
• TPC-H SF50: dataset/tpch/tbl-sf50, dataset/tpch/parquet-sf50
• TPC-H SF100: dataset/tpch/tbl-sf100, dataset/tpch/parquet-sf100

We have two sets of queries using TPC-H datasets: vanilla and tpc-h.

TPC-DS Dataset

First, generating the original tables (dat format) using TPC-DS tools, and simply running duckdb_tpcds_data.py can convert the table files to parquet or csv format using the following command,

cd tpcds
python3 duckdb_tpcds_data.py -d ../dataset/dat-sf1 -f parquet -rgs 10000

We have several datasets:

• TPC-DS SF-0.01 (Tiny): dataset/tpcds/dat-tiny, dataset/tpcds/parquet-tiny
• TPC-DS SF-0.1 (Small): dataset/tpcds/dat-small, dataset/tpcds/parquet-small
• TPC-DS SF1: dataset/tpcds/dat-sf1, dataset/tpcds/parquet-sf1
• TPC-DS SF10: dataset/tpcds/dat-sf10, dataset/tpcds/parquet-sf10
• TPC-DS SF50: dataset/tpcds/dat-sf50, dataset/tpcds/parquet-sf50
• TPC-DS SF100: dataset/tpcds/dat-sf100, dataset/tpcds/parquet-sf100

Experiments

Benchmark for Pipeline-level Suspension and Resumption

# choose the benchmark: tpch, or tpcds
bm=[tpch, tpcds]
# make sure in <bm> folder
cd <bm>
# the queries should be in <bm/queries>, the name the argument for "-q" option
# for example, q1-q22 in tpch, and q1-q99 in tpcds
# run q1 based on xxx.db and the dataset from parquet-tiny using 2 threads
python3 ratchet_<bm>.py -q q1 -d xxx.db -df ../dataset/<bm>/parquet-sf10 -td 2

# run q1 with a suspension point determined by time window [st, se] with uniform distribution, and serialize into single file
python3 ratchet_<bm>.py -q q1 -d xxx.db -df ../dataset/<bm>/parquet-sf10 -td 2 -s -st 0 -se 0 -sl yyy.ratchet 
# run q1 with resumption using a single file
python3 ratchet_<bm>.py -q q1 -d xxx.db -df ../dataset/<bm>/parquet-sf10 -td 2 -r -rl yyy.ratchet 

# run q1 with a suspension point determined by time window [st, se] with uniform distribution
# and serialize into multiple files (will generate part-*.ratchet in demo folder)
python3 ratchet_<bm>.py -q q1 -d xxx.db -df ../dataset/<bm>/parquet-sf10 -td 2 -s -st 0 -se 0 -sl ./ -psr
# run q1 with resumption using multiple files (will use all part-*.ratchet in the demo folder)
python3 ratchet_<bm>.py -q q1 -d xxx.db -df ../dataset/<bm>/parquet-sf10 -td 2 -r -rl ./ -psr

Benchmark for Process-level Suspension and Resumption

We also benchmark the performance of suspending and resuming queries at the process level. More details can be found here.

Benchmark for Suspension and Resumption with Cost Model

MISC

Ubuntu Disk Commands

# Test the speed of cached read and buffered write
sudo hdparm -Tt <disk name, such as /dev/nvme0n1p3>

# Test the speed of random write 
sudo fio -ioengine=libaio -bs=4k -direct=1 -thread -rw=randwrite -filename=/dev/nvme0n1p3 -name="BS 4k randwrite test" -iodepth=16 -runtime=30
# Test the speed of rand read 
sudo fio -ioengine=libaio -bs=4k -direct=1 -thread -rw=randread -filename=/dev/nvme0n1p3 -name="BS 4k randread test" -iodepth=16 -runtime=30

With df / -h, you will know the disk where your root filesystem and OS (e.g., Ubuntu) is located.

$ df / -h 
Filesystem      Size  Used Avail Use% Mounted on
/dev/nvme0n1p3  449G  145G  282G  34% /

lshw will give more details about the disks or SSDs

$ sudo lshw -short -C disk
H/W path         Device          Class          Description
===========================================================
/0/100/1d/0/0    hwmon1          disk           NVMe disk
/0/100/1d/0/2    /dev/ng0n1      disk           NVMe disk
/0/100/1d/0/1    /dev/nvme0n1    disk           500GB NVMe disk

lsblk can further show something such as rotation, type, and mountpoints.

$ lsblk
NAME          MAJ:MIN RM   SIZE RO TYPE  MOUNTPOINTS
zram0         252:0    0    16G  0 disk  [SWAP]
nvme0n1       259:0    0 465.8G  0 disk  
├─nvme0n1p1   259:1    0  1022M  0 part  /boot/efi
├─nvme0n1p2   259:2    0     4G  0 part  /recovery
├─nvme0n1p3   259:3    0 456.8G  0 part  /
└─nvme0n1p4   259:4    0     4G  0 part  
  └─cryptswap 253:0    0     4G  0 crypt [SWAP]