krzysztof/landing_page
1
0
Fork 0
mirror of https://github.com/qdrant/landing_page.git synced 2026-09-28 23:48:31 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #1056 from qdrant/chore/fix-units

Browse Source 
Use proper units (gigabytes and megabytes, not gigabits and megabits)
This commit is contained in:
David Myriel
2024-07-31 18:59:23 -07:00
committed by GitHub
parent fc62618a52 9597f68f07
commit 6d1f7e9aec
4 changed files with 11 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
qdrant-landing/content/articles/bm42.md
+1 -1
View File
@@ -340,7 +340,7 @@ As you can see, it has pretty short texts, there are not much of the statistics
After encoding with BM42, the average vector size is only **5.6 elements per document**.
With `datatype: uint8` available in Qdrant, the total size of the sparse vector index is about **13Mb** for ~530k documents.
With `datatype: uint8` available in Qdrant, the total size of the sparse vector index is about **13MB** for ~530k documents.
As a reference point, we use:
qdrant-landing/content/articles/dedicated-service.md
+1 -1
View File
@@ -99,7 +99,7 @@ Assume we have a database with 1 million records.
This is a small database by modern standards of any relational database.
You can probably use the smallest free tier of any cloud provider to host it.
But if we want to use this database for vector search, 1 million OpenAI `text-embedding-ada-002` embeddings will take **~6Gb of RAM** (sic!).
But if we want to use this database for vector search, 1 million OpenAI `text-embedding-ada-002` embeddings will take **~6GB of RAM** (sic!).
As you can see, the vector search use case completely overwhelmed the main database resource requirements.
In practice, this means that your main database becomes burdened with high memory requirements and can not scale efficiently, limited by the size of a single machine.
qdrant-landing/content/articles/memory-consumption.md
+4 -4
View File
@@ -29,7 +29,7 @@ Introduction:
2. Process may not free deallocated memory.
3. Process might be forked and memory is shared between processes.
3. Process may use disk cache.
3. As a result, if you see `10Gb` memory consumption in `htop`, it doesn't mean that your process actually needs `10Gb` of RAM to work.
3. As a result, if you see `10GB` memory consumption in `htop`, it doesn't mean that your process actually needs `10GB` of RAM to work.
-->
# Mastering RAM Measurement and Memory Optimization in Qdrant: A Comprehensive Guide
@@ -120,7 +120,7 @@ All in memory:
1152mb - out of memory
1200mb - 794.72it/s
Conclusion: about 1.2Gb is needed to serve ~1 million vectors, no speed degradation with limiting memory above 1.2Gb
Conclusion: about 1.2GB is needed to serve ~1 million vectors, no speed degradation with limiting memory above 1.2GB
MMAP for vectors:
@@ -164,8 +164,8 @@ We tried using different amounts of memory, ranging from 1512mb to 1024mb, and m
| 1024mb | out of memory |
We found that 1152Mb memory limit resulted in our system running out of memory, but using 1512mb, 1256mb, and 1200mb of memory resulted in our system being able to handle around 780 RPS.
This suggests that about 1.2Gb of memory is needed to serve around 1 million vectors, and there is no speed degradation when limiting memory usage above 1.2Gb.
We found that 1152MB memory limit resulted in our system running out of memory, but using 1512mb, 1256mb, and 1200mb of memory resulted in our system being able to handle around 780 RPS.
This suggests that about 1.2GB of memory is needed to serve around 1 million vectors, and there is no speed degradation when limiting memory usage above 1.2GB.
### Vectors stored using MMAP
qdrant-landing/content/articles/scalar-quantization.md
+5 -5
View File
@@ -268,16 +268,16 @@ We used a machine with a very slow network-mounted disk and tested the following
| Setup | RPS | Precision |
|-----------------------------|------|-----------|
| 4.5Gb memory | 600 | 0.99 |
| 4.5Gb memory + SQ + rescore | 1000 | 0.989 |
| 4.5GB memory | 600 | 0.99 |
| 4.5GB memory + SQ + rescore | 1000 | 0.989 |
And another group with more strict memory limits:
| Setup | RPS | Precision |
|------------------------------|------|-----------|
| 2Gb memory | 2 | 0.99 |
| 2Gb memory + SQ + rescore | 30 | 0.989 |
| 2Gb memory + SQ + no rescore | 1200 | 0.974 |
| 2GB memory | 2 | 0.99 |
| 2GB memory + SQ + rescore | 30 | 0.989 |
| 2GB memory + SQ + no rescore | 1200 | 0.974 |
In those experiments, throughput was mainly defined by the number of disk reads, and quantization efficiently reduces it by allowing more vectors in RAM.
Read more about on-disk storage in Qdrant and how we measure its performance in our article: [Minimal RAM you need to serve a million vectors