On-disk Index
This article introduces an on-disk indexing algorithm named DiskANN. Based on Vamana graphs, DiskANN powers efficient searches within large datasets.
To improve query performance, you can specify an index type for each vector field.
Prerequisites
To use DiskANN, note that
- DiskANN is enabled by default. If you prefer in-memory index over on-disk index, you are advised to disable this feature for a better performance.
- To disable it, you can change
queryNode.enableDisktofalsein your milvus configuration file. - To enable it again, you can set
queryNode.enableDisktotrue.
- To disable it, you can change
- The Milvus instance runs on Ubuntu 18.04.6 or a later release.
- The Milvus data path should be mounted to an NVMe SSD for full performance:
- For a Milvus Standalone instance, the data path should be /var/lib/milvus/data in the container where the instance runs.
- For a Milvus Cluster instance, the data path should be /var/lib/milvus/data in the containers where the QueryNodes and IndexNodes run.
Limits
To use DiskANN, ensure that you
- Use only float vectors with at least 32 dimensions in your data.
- Use only Euclidean Distance (L2) or Inner Product (IP) to measure the distance between vectors.
Index and search settings
Index building parameters
When building a DiskANN index, use
DISKANNas the index type. No index parameters are necessary.Search parameters
Parameter Description Range search_listSize of the candidate list, a larger size offers a higher recall rate with degraded performance. [k, min( 10 * k, 65535)] for k > 20
[k, 200] for k <= 20
DiskANN-related Milvus configurations
DiskANN is tunable. You can modify DiskANN-related parameters in ${MILVUS_ROOT_PATH}/configs/milvus.yaml to improve its performance.
...
DiskIndex:
MaxDegree: 56
SearchListSize: 100
PQCodeBugetGBRatio: 0.125
BuildNumThreadsRatio: 1.0
SearchCacheBudgetGBRatio: 0.125
LoadNumThreadRatio: 8.0
BeamWidthRatio: 4.0
...
| Parameter | Description | Value Range | Default Value |
|---|---|---|---|
MaxDegree | Maximum degree of the Vamana graph. A larger value offers a higher recall rate but increases the size of and time to build the index. | [1, 512] | 56 |
SearchListSize | Size of the candidate list. A larger value increases the time spent on building the index but offers a higher recall rate. Set it to a value smaller than MaxDegree unless you need to reduce the index-building time. | [1, ∞] | 100 |
PQCodeBugetGBRatio | Size limit on the PQ code. A larger value offers a higher recall rate but increases memory usage. | (0.0, 0.25] | 0.125 |
BuildNumThreadsRatio | Ratio between the number of threads used to build the index and the number of CPUs. | [1.0, 128.0 / CPU number] | 1.0 |
SearchCacheBudgetGBRatio | Ratio of cached node numbers to raw data. A larger value improves index-building performance with increased memory usage. | [0.0, 0.3) | 0.10 |
LoadNumThreadRatio | Ratio between the number of threads used to load index/search and the number of CPUs. For details, refer to the first item in References and Facts. | [1, 65536 / 32 / CPU number] | 8.0 |
BeamWidthRatio | Ratio between the maximum number of IO requests per search iteration and CPU number. | [1, max(128 / CPU number, 16)] | 4.0 |
Troubleshooting
How to deal with the
io_setup() failed; returned -11, errno=11:Resource temporarily unavailableerror?The Linux kernel provides the Asynchronous non-blocking I/O (AIO) feature that allows a process to initiate multiple I/O operations simultaneously without having to wait for any of them to complete. This helps boost performance for applications that can overlap processing and I/O.
The performance can be tuned using the
/proc/sys/fs/aio-max-nrvirtual file in the proc file system. Theaio-max-nrparameter determines the maximum number of allowable concurrent requests.The
aio-max-nrdefaults to65535, you can set it up to10485760.