Skip to content

The fastest ACID-transactional persisted Key-Value store designed as modified LSM-Tree for NVMe block-devices with GPU-acceleration and SPDK to bypass the Linux kernel

Notifications You must be signed in to change notification settings

unum-cloud/udisk

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 

Repository files navigation

UDisk

The fastest ACID-transactional persisted Key-Value store designed for NVMe block-devices with GPU-acceleration and SPDK to bypass the Linux kernel.

Configuration Parameters

Most important parameters:

  • directory: Where to keep the configs, schemas, data, and logs? The boot drive?
  • memory_limit: How much RAM to use? 80% of it?
  • value_max_size: How large the values can get?

Tuning performance?

  • directories_data: Want to separate large data and put it on some RAID array?
  • gpu_devices: Do you a have a GPU we can borrow for acceleration?

Versioning

"version": 1

Disk IO

Path where there database would store metadata:

"directory": "/etc/udisk/"

Paths where the database would store data, with tuples containing the path and a limit:

"data_directories": [
    {
        "path": "/mnt/md0/ukv/",
        "max_size": 1e12,
    },
    {
        "path": "/mnt/md1/ukv/",
        "max_size": 1e12,
    }
]

Choose the underlying IO mechanism:

"io_mechanism": "posix"
"io_mechanism": "pulling"
"io_mechanism": "polling"

For asynchronous IO mechanisms you can set the queue depth:

"io_queue_depth": 4096

Random Access Memory

Limit the total amount of RAM available to the database:

"memory_limit": 1000000000

Limit the amount of memory to be used for caching. Greatly affects the performance of single reads. Doesn't affect the performance of batch reads.

"cache_limit": 100000000

Limit the maximum number of entries to be collected in-memory (and in the Write Ahead Log), before flushing them into files.

"write_buffer_max_elements": 100000

Limit the maximum combined size of entries to be collected in-memory (and in the Write Ahead Log), before flushing them into files.

"write_buffer_max_bytes": 100000000

Hardware Acceleration

Threads to be used for compactions, non-blocking garbage collection, and polling mechanisms. For optimal performance with high-end NVMe SSDs we suggest reserving 2 or 3 threads per SSD.

"threads": 1

Define, how many threads may be accessing the database concurrently:

"concurrency_limit": 64

Optional array of integer identifiers for the GPUs to be used for background compactions:

"gpu_devices": "all"
"gpu_devices": [-1]
"gpu_devices": [0, 1, 2, 3]

The amount of Video RAM available to UDisk on the GPU:

"gpu_memory_limit": 1000000000

Structure

Define, how large the biggest value can be:

"value_max_size": 4096

In some rare cases, you may have all of your values of the same size:

"value_size": 0

Define, how much larger every layer of the Log-Structured-Merge Tree can get:

"level_enlarge_factor": 4

How large the top-level of the Log-Structured-Merge Tree should be?

"first_level_max_bytes": 1000000000

Transactions

The maximum number of entries to be updated with a single transaction:

"transaction_max_elements": 100000

The maximum capacity of a single transaction, limiting the size of all updates in it:

"transaction_max_bytes": 100000000

Full Example

{
    "version": 1,
    "threads": 1,
    "memory_limit": 100000000000,
    "gpu_devices": [-1],
    "gpu_memory_limit": 1000000000,
    "directory": "./tmp/",
    "data_directories": [
        {
            "path": "/mnt/disk1/",
            "max_size": 100000000000
        },
        {
            "path": "/mnt/disk2/",
            "max_size": 100000000000
        }
    ],
    "io_mechanism": "pulling",
    "io_queue_depth": 4096,
    "transaction_max_elements": 1000,
    "transaction_max_bytes": 13421773,
    "concurrency_limit": 128,
    "cache_limit": 134217728,
    "write_buffer_max_elements": 100000,
    "write_buffer_max_bytes": 134217728,
    "first_level_max_bytes": 4831838208,
    "level_enlarge_factor": 4,
    "value_size": 0,
    "value_max_size": 4194304
}

About

The fastest ACID-transactional persisted Key-Value store designed as modified LSM-Tree for NVMe block-devices with GPU-acceleration and SPDK to bypass the Linux kernel

Topics

Resources

Stars

Watchers

Forks