Resource Metrics

Overview

Details of specific measurements used to assess and quantify the resource utilization of LILT applications. These metrics provide insights into the performance, capacity, and scalability requirements of the applications. They include parameters such as CPU utilization, memory consumption, disk size and concurrent user sessions. By establishing baseline measurements and setting thresholds, usage-sizing metrics help identify normal operating levels and define boundaries for resource usage. These metrics enable proactive capacity planning, resource optimization, and the ability to detect anomalies or performance bottlenecks. The table below is meant to be a reference guide, and not an exact match to your individual environment, as each installation may choose to enable or disable specific services.

Usage-Sizing Metrics

App Name	Min Recommended vCPUs	Max Recommended vCPUs	Min Recommended Memory (GB)	Max Recommended Memory (GB)	GPU Requirements
Web Application Services
Front	0.5	2	4	8
Beehive	0.1	4	4	4
Core Translation Services
Core-api	0.01	0.1	0.256	0.256
Converter	1	12	26	28
QA	0.5	1	3	4
Linguist	0.5	1	1	2
Search	0.05	n/a	21	22
TM	4	12	30	32
TB	1	n/a	30	30
Batch-TB	2	7	22	30
Indexer	0.05	n/a	21	22
Watchdog	0.1	n/a	2	3
Segment	0.5	2	4	4
File-translation	3	6	14	15
Job	0.5	1	2	3.5
Tag	0.5	1	2	3.5
Auditlog	0.5	1	2	3.5
Assignment	1	2	4	4
Workflow	1	2	2	2
File-job	2	4	16	16
Memory	1	2	4	4.5
Notification	0.5	n/a	2	4
Auth	0.01	n/a	0.5	1
Neural/ML Services
Translatev4****	0.7	0.7	32	32	2 GPUs
Updatev4	0.6	0.6	38	38
Langid	0.5	3	5	5
Update-managerv4	0.4	0.4	2.5	2.5
Routing	0.1	0.35	1.5	1.5
Batchv4	0.25	0.5	2	2
Batch-worker-gpuv4****	1.2	1.2	18	19	1 GPU
Batch-worker-cpu-3pmt	5	5	18	18
Llama-vllm-inference****	5	14	100	100	4 GPUs
LLM-inference-whisper****	1	n/a	8	8	1 GPU
Gemma-vllm-inference****	12	12	80	80	2 GPUs
Automqm	0.4	0.4	3	3
Alignment	0.7	0.7	13	13
Tag-projection	0.8	0.8	6	6
NNcache	0.5	0.5	5	5
VMF (Vector Manipulation)	2	4	8	8
Rayman****	6	12	24	72	1 GPU (+1 optional)
Connectors Platform
Connectors-ingressgateway	0.1	2	0.128	1
Connectors-builder	n/a	n/a	n/a	n/a
Infrastructure Services
nginx-ingress	0.1	n/a	0.128	n/a
istiod	0.5	0.5	2	2
istio-ingressgateway	0.1	2	0.128	1
Qdrant***	4	n/a	12	12
Prometheus	n/a	0.5	n/a	2
Grafana	n/a	0.5	n/a	0.5
Data Stores
Redis	0.2	0.6	4	10
RabbitMQ	1	2	4	4
MySQL	0.5	n/a	4	8
MongoDB	0.5	0.5	0.5	0.5
MinIO	0.1	0.5	1	1.5
ClickHouse	0.5	n/a	1	16
Elasticsearch***	1	2	4	4
Total*	49.77 vCPUs	147.25 vCPUs	633.884 GB RAM	748.756 GB RAM	11 GPUs
Total Disk Requirements*	1 TB minimum

*The following specifications are recommended for installing LILT on bare-metal hardware to support the given applications. It is important to note that these specifications may need to be adjusted based on the specific load requirements.

**To efficiently handle parallel batch requests, it is recommended to have the same number of GPUs as the number of required concurrent batch requests. This enables the system to effectively distribute the workload and process multiple requests simultaneously, leveraging the processing power of each GPU to improve throughput.

*** These values may be doubled when multiple replicas are configured

**** These pods require GPU nodes for operation

Rayman is the neural translation service on Ray (Ray Serve / KubeRay). It runs on a single GPU by default; the neural v4.0 pipeline that provides optional Uzbek and Chechen language support (replacing the former Emma model) requires one additional GPU. vCPU and memory ranges reflect the combined Ray head, application, and model worker groups.

n/a marked services have no specified limit in the configuration and can use available resources on the node. Services marked with n/a for both min and max have no explicit resource specifications defined in the helm charts and will use default values.

​Overview

​Usage-Sizing Metrics