Skip to content

CONFIGURATION.md

Latest commit

 

History

History
517 lines (418 loc) · 28.5 KB

CONFIGURATION.md

File metadata and controls

517 lines (418 loc) · 28.5 KB

Configuration Guide

This document provides detailed configuration documentation for TURBO. For a quick start, see the main README.md.

Overview

The system uses four YAML configuration files:

File Used by Purpose
config/client_config.yaml Python client processes Camera streams, clients, bandwidth allocator, ping handler, plotter
config/server_config_gcloud.yaml Python server processes Model servers and GPU assignments
config/quic_config_client.yaml Rust QUIC client binary QUIC client transport, logging, junk service
config/quic_config_gcloud.yaml Rust QUIC server binary QUIC server transport, logging

The client and server QUIC configs share the same schema but may have different values (e.g., different junk_tx_loop_interval_ms for local vs cloud).

Naming Conventions

The config files make heavy use of YAML anchors (&anchor-name) and references (*anchor-name) to avoid duplication. For example, SLO_TIMEOUT: &slo-timeout 200 defines the value once, and each per-service entry references it with SLO_TIMEOUT_MS: *slo-timeout.

IPC Socket Path Convention

ZMQ IPC socket names are stored as bare names (e.g., service1-camera-socket) in all YAML config files. The Python orchestrators (client_main.py, server_main.py) automatically resolve them to full ipc:// paths using the zmq_dir directory specified in the config. For example, a bare name service1-camera-socket with zmq_dir: /home/user/experiment-out/zmq becomes ipc:///home/user/experiment-out/zmq/service1-camera-socket. See IPC.md for the full socket reference.

Auto-Generated Output Directories

Save directories for logs (client_savedir, camera_savedir, ping_savedir, server_log_savedir) are not specified in the YAML configs. Instead, each orchestrator creates a timestamped run directory at startup under experiment_output_dir and injects the appropriate save path into each component's config. For example, client_main.py creates experiment_output_dir/client_main_2024-01-15_10-30-00/client/ and sets that as the save directory for all client-side components. The QUIC binaries similarly create timestamped subdirectories (e.g., experiment_output_dir/quic_client_2024-01-15_10-30-00/quic-client-out/).

1. Client Configuration — config/client_config.yaml

Configures the client-side (AV) components: camera streams, perception service clients, bandwidth allocator, ping handler, web dashboard, and diagnostic plotter.

Global Parameters

logging_config_filepath: /path/to/turbo/config/logging_config.yaml

experiment_output_dir: /path/to/experiment-out
client_subdir: client
zmq_dir: /path/to/experiment-out/zmq

DST_IP: &dst-ip <YOUR_SERVER_IP>    # Cloud server IP for ICMP pings
SLO_TIMEOUT: &slo-timeout 200       # Client-side SLO timeout (ms), shared via anchor
QUIC_SHM_SIZE: &quic-shm-size 50000000  # Shared memory region size (bytes, ~50 MB)
MAX_LOG_ENTRIES: &max-log-entries 100    # Max records in memory before spilling to Parquet
Parameter Must customize? Description
logging_config_filepath Yes Absolute path to the Python logging config file
experiment_output_dir Yes Base output directory. client_main.py creates a timestamped subdirectory here (e.g., experiment-out/client_main_2024-01-15_10-30-00/) for each run
client_subdir No Subdirectory name within the timestamped run directory for client Parquet logs. Default client
zmq_dir Yes Directory for ZMQ IPC socket files. Must match across client, server, and QUIC configs. Created automatically if it doesn't exist
DST_IP Yes Public IP of your cloud server. Used by PingHandler for RTT measurement
SLO_TIMEOUT Maybe Service-level objective timeout in ms. Frames exceeding this are dropped. Default 200 is suitable for most setups
QUIC_SHM_SIZE No Size of each POSIX shared memory region in bytes. 50000000 (50 MB) is sufficient for HD frames
MAX_LOG_ENTRIES No Number of log records buffered in memory before flushing to Parquet. Higher = fewer I/O flushes, more memory

Web Dashboard Configuration

web_dashboard_config:
  refresh_rate_seconds: 6
  plotting_loop_sleep_seconds: 2
Parameter Must customize? Description
refresh_rate_seconds No How often the web dashboard refreshes plots (seconds)
plotting_loop_sleep_seconds No Sleep interval for the plotting loop (seconds)

Model Image Sizes

model_image_size_map: &model_imagesizes
  tf_efficientdet_d1: [640, 640]
  tf_efficientdet_d2: [768, 768]
  tf_efficientdet_d4: [1024, 1024]
  tf_efficientdet_d6: [1280, 1280]
  tf_efficientdet_d7x: [1536, 1536]

Maps each EfficientDet variant to its native input resolution. Do not change unless using custom-trained models with different input sizes.

Per-Service Client Configuration (main_client_config_list)

One entry per perception service. Add or remove entries to match your number of cameras/services.

main_client_config_list:
  - service_id: 1
    max_entries: *max-log-entries
    thread_concurrency: 10
    camera_bidirectional_zmq_sockname: service1-camera-socket
    camera_stream_shmem_filename: service1-camera-shmem
    bandwidth_allocation_incoming_zmq_sockname: main-client-1-bw-subscriber
    quic_rcv_zmq_sockname: car-server-outgoing-1
    quic_snd_zmq_sockname: car-server-incoming-1
    outgoing_zmq_diagnostic_sockname: car-client-diagnostics
    camera_np_size: [1080, 1920, 3]
    model_name_imagesize_map: *model_imagesizes
    zmq_kill_switch_sockname: client-kill-1-switch
    quic_snd_shm_filename: client-service1-incoming-shm
    quic_rcv_shm_filename: client-service1-outgoing-shm
    quic_shm_size: *quic-shm-size
    SLO_TIMEOUT_MS: *slo-timeout

Note: client_savedir is not specified in the YAML. It is automatically injected by client_main.py using the timestamped run directory.

Parameter Must customize? Description
service_id Yes (if adding/removing services) Unique integer ID for this service. Must match across client, server, and QUIC configs
max_entries No Log buffer size before spilling to Parquet (uses global anchor)
thread_concurrency No Number of threads in the client's thread pool. Default 10 is suitable for most setups
camera_bidirectional_zmq_sockname No Bare ZMQ socket name for the CameraDataStream. Auto-resolved to ipc://<zmq_dir>/<name> by client_main.py
camera_stream_shmem_filename No POSIX SHM region name for raw camera frames. Must match the corresponding camera stream entry
bandwidth_allocation_incoming_zmq_sockname No Bare ZMQ socket name for receiving bandwidth allocation updates from the BandwidthAllocator
quic_rcv_zmq_sockname No Bare ZMQ socket name for sending compressed images to the QUIC client
quic_snd_zmq_sockname No Bare ZMQ socket name for receiving inference results from the QUIC client
outgoing_zmq_diagnostic_sockname No Bare ZMQ socket name for diagnostic messages to the plotter. All services share the same name
camera_np_size Maybe Camera frame dimensions as [height, width, channels]. Change if your cameras produce a different resolution
model_name_imagesize_map No Reference to the global model image size map (uses anchor)
zmq_kill_switch_sockname No Bare ZMQ socket name for receiving graceful shutdown signals from client_main.py
quic_snd_shm_filename No POSIX SHM region name for outgoing images to QUIC
quic_rcv_shm_filename No POSIX SHM region name for incoming results from QUIC
quic_shm_size No Size of the QUIC shared memory regions (uses global anchor)
SLO_TIMEOUT_MS No Per-service SLO deadline in ms (uses global anchor)

When adding a new service, duplicate an existing entry, increment service_id, and update all socket names and SHM filenames to use the new service number (e.g., replace -1 with -4).

Bandwidth Allocator Configuration (bandwidth_allocator_config)

bandwidth_allocator_config:
  service_id_list: [1, 2, 3]
  t_SLO: 150
  parquet_eval_dir: /path/to/full-eval
  model_info_csv_path: /path/to/turbo/experiment_model_info.csv
  outgoing_zmq_diagnostic_sockname: car-client-diagnostics
  outgoing_zmq_client_socknames:
    - main-client-1-bw-subscriber
    - main-client-2-bw-subscriber
    - main-client-3-bw-subscriber
  bidirectional_zmq_quic_sockname: car-server-bw-service
  zmq_kill_switch_sockname: bandwidth-allocator-kill-switch
  bidirectional_zmq_ping_handler_sockname: ping-handler
Parameter Must customize? Description
service_id_list Yes (if adding/removing services) List of active service IDs to allocate bandwidth across. Must not include the junk service
t_SLO Maybe Latency SLO constraint used by the LP solver (ms). Should be slightly tighter than SLO_TIMEOUT to account for processing overhead
parquet_eval_dir Yes Path to directory containing pre-computed utility curve Parquet files. These are generated offline and encode the bandwidth-to-accuracy mapping for each model configuration
model_info_csv_path Yes Path to experiment_model_info.csv, which maps model configuration strings to transport size (Mb) and runtime (ms)
outgoing_zmq_diagnostic_sockname No Bare ZMQ socket name for allocation diagnostics. Should match the diagnostic socket used by clients
outgoing_zmq_client_socknames No List of bare ZMQ socket names, one per service. Each must match the corresponding client's bandwidth_allocation_incoming_zmq_sockname
bidirectional_zmq_quic_sockname No Bare ZMQ socket name for receiving bandwidth/RTT updates from the QUIC client
zmq_kill_switch_sockname No Bare ZMQ socket name for graceful shutdown
bidirectional_zmq_ping_handler_sockname No Bare ZMQ socket name for querying RTT from the PingHandler

Camera Stream Configuration (camera_stream_config_list)

One entry per USB camera. Must have a matching entry in main_client_config_list.

camera_stream_config_list:
  - camera_id: 1                     # Must match service_id
    usb_id: 0                        # USB device index
    max_entries: *max-log-entries
    thread_concurrency: 10
    bidirectional_zmq_sockname: service1-camera-socket
    camera_stream_shmem_filename: service1-camera-shmem
    shmem_buf_size: *quic-shm-size
    camera_np_size: [1080, 1920, 3]
    zmq_kill_switch_sockname: camera-kill-1-switch
    mock_camera_image_path: mock_webcam_image.jpg  # set to null for live camera

Note: camera_savedir is not specified in the YAML. It is automatically injected by client_main.py using the timestamped run directory.

Parameter Must customize? Description
camera_id Yes (if adding/removing services) Must match the service_id of the corresponding client entry
usb_id Yes USB camera device index for cv2.VideoCapture. Run ls /dev/video* to identify your cameras. These indices vary by system
max_entries No Log buffer size (uses global anchor)
thread_concurrency No Thread pool size for camera operations
bidirectional_zmq_sockname No Bare ZMQ socket name for client communication. Must match the corresponding client's camera_bidirectional_zmq_sockname
camera_stream_shmem_filename No SHM region name. Must match the corresponding client's camera_stream_shmem_filename
shmem_buf_size No SHM buffer size in bytes (uses global anchor)
camera_np_size Maybe Camera frame dimensions [height, width, channels]. Must match the client entry
zmq_kill_switch_sockname No Bare ZMQ socket name for graceful shutdown
mock_camera_image_path Maybe Set to a file path (e.g., mock_webcam_image.jpg) to use a static image instead of a live webcam. Useful for testing without hardware. Set to null for live camera

Ping Handler Configuration (ping_handler_config)

ping_handler_config:
  dst_ip: *dst-ip
  max_entries: 100
  thread_concurrency: 5
  bidirectional_zmq_sockname: ping-handler
  zmq_kill_switch_sockname: ping-handler-kill-switch

Note: ping_savedir is not specified in the YAML. It is automatically injected by client_main.py using the timestamped run directory.

Parameter Must customize? Description
dst_ip Yes IP address to ping for RTT measurement. Should be your cloud server's public IP (uses global DST_IP anchor)
max_entries No Log buffer size
thread_concurrency No Thread pool size
bidirectional_zmq_sockname No Bare ZMQ socket name. Must match the bandwidth allocator's bidirectional_zmq_ping_handler_sockname
zmq_kill_switch_sockname No Bare ZMQ socket name for graceful shutdown

Diagnostic Plotter Configuration (main_plotter_config)

Configures the real-time matplotlib diagnostic plots. This section is optional — the plotter is started separately and can be omitted if you only use the web dashboard.

main_plotter_config:
  plotting_loop_sleep_seconds: 2
  zmq_incoming_diagnostic_name: car-client-diagnostics
  bandwidth_allocation_plot_config:
    service_id_list: [1, 2, 3]
    window_size_x: 40               # X-axis window size (seconds of data shown)
    bw_min_y: -10                    # Bandwidth plot Y-axis minimum
    bw_max_y: null                   # Bandwidth plot Y-axis maximum (null = auto)
    utility_min_y: -0.1              # Utility plot Y-axis minimum
    utility_max_y: 1.1               # Utility plot Y-axis maximum
    # Tick mark spacing (major/minor) for each subplot:
    bw_x_major_loc: 15
    bw_x_minor_loc: 5
    bw_y_major_loc: 100
    bw_y_minor_loc: 20
    utility_x_major_loc: 15
    utility_x_minor_loc: 5
    utility_y_major_loc: 0.2
    utility_y_minor_loc: 0.1

  service_status_plot_config:        # One entry per service
    - service_id: 1
      window_size_x: 40
      cnt_min_y: -3                  # Request count Y-axis minimum
      cnt_max_y: null
      rate_min_y: -0.5               # Success rate Y-axis minimum
      rate_max_y: 1.05
      # Tick mark spacing for count and rate subplots:
      cnt_x_major_loc: 15
      cnt_x_minor_loc: 5
      cnt_y_major_loc: 30
      cnt_y_minor_loc: 10
      rate_x_major_loc: 15
      rate_x_minor_loc: 5
      rate_y_major_loc: 0.2
      rate_y_minor_loc: 0.05

  service_utilization_plot_config:   # One entry per service (including junk)
    - service_id: 1
      window_size_x: 40
      min_y: -10                     # Utilization Y-axis minimum
      max_y: null
      x_major_loc: 15
      x_minor_loc: 5
      y_major_loc: 100
      y_minor_loc: 20
Parameter Must customize? Description
plotting_loop_sleep_seconds No Sleep interval between plot updates
zmq_incoming_diagnostic_name No Bare ZMQ socket name for receiving diagnostics from clients, bandwidth allocator, and QUIC. Must match outgoing_zmq_diagnostic_sockname used by other components
bandwidth_allocation_plot_config No Plot layout for the bandwidth allocation overview. Adjust window_size_x to show more or less history
service_status_plot_config Yes (if adding/removing services) One entry per service. Adjust to match your service_id_list
service_utilization_plot_config Yes (if adding/removing services) One entry per service including junk service (service 4). Adjust to match your services list

2. Server Configuration — config/server_config_gcloud.yaml

Configures the server-side (cloud) components: model servers and GPU assignments.

Global Parameters

logging_config_filepath: /path/to/turbo/config/logging_config.yaml

experiment_output_dir: /path/to/experiment-out
server_subdir: server
zmq_dir: /path/to/experiment-out/zmq

MAX_LOG_ENTRIES: &max-log-entries 100
SHM_FILESIZE: &shm-filesize 50000000
Parameter Must customize? Description
logging_config_filepath Yes Absolute path to the Python logging config file
experiment_output_dir Yes Base output directory. server_main.py creates a timestamped subdirectory here (e.g., experiment-out/server_main_2024-01-15_10-30-00/) for each run
server_subdir No Subdirectory name within the timestamped run directory for server Parquet logs. Default server
zmq_dir Yes Directory for ZMQ IPC socket files. Must match across client, server, and QUIC configs. Created automatically if it doesn't exist
MAX_LOG_ENTRIES No Log buffer size before spilling to Parquet
SHM_FILESIZE No Shared memory region size in bytes (~50 MB)

Model Metadata (server_model_list)

Defines all available EfficientDet model checkpoints. These are shared across all model servers via YAML anchor.

server_model_list: &model-list-ref
  - checkpoint_path: /path/to/av-models/tf_efficientdet_d2-waymo-open-dataset/version_2/checkpoints/epoch=9-step=419700.ckpt
    num_classes: 5
    image_size: [768, 768]
    base_model: "tf_efficientdet_d2"
  - checkpoint_path: /path/to/av-models/tf_efficientdet_d4-waymo-open-dataset/version_0/checkpoints/epoch=9-step=839400.ckpt
    num_classes: 5
    image_size: [1024, 1024]
    base_model: "tf_efficientdet_d4"
  - checkpoint_path: /path/to/av-models/tf_efficientdet_d6-waymo-open-dataset/version_2/checkpoints/epoch=9-step=3357600.ckpt
    num_classes: 5
    image_size: [1280, 1280]
    base_model: "tf_efficientdet_d6"
  - checkpoint_path: /path/to/av-models/tf_efficientdet_d7x-waymo-open-dataset/version_1/checkpoints/epoch=8-step=1477071.ckpt
    num_classes: 5
    image_size: [1536, 1536]
    base_model: "tf_efficientdet_d7x"
Parameter Must customize? Description
checkpoint_path Yes Absolute path to each PyTorch Lightning .ckpt file. Update after downloading models (see MODELS.md)
num_classes No Number of detection classes. 5 for the provided Waymo-trained models
image_size No Native input resolution [height, width] for each model variant. Must match the model_image_size_map in the client config
base_model No Model variant name matching the effdet library naming convention

Per-Service Server Configuration (server_config_list)

One entry per perception service. Add or remove entries to match your number of services.

server_config_list:
  - service_id: 1
    max_entries: *max-log-entries
    model_metadata_list: *model-list-ref
    device: "cuda:0"
    incoming_zmq_sockname: remote-server-outgoing-1
    incoming_shm_filename: server-service1-outgoing-shm
    outgoing_zmq_sockname: remote-server-incoming-1
    outgoing_shm_filename: server-service1-incoming-shm
    thread_concurrency: 10
    shm_filesize: *shm-filesize
    zmq_kill_switch_sockname: remote-server-kill-switch-1
    mock_inference_output_path: null
    mock_model_latency_csv_path: null

Note: server_log_savedir is not specified in the YAML. It is automatically injected by server_main.py using the timestamped run directory.

Parameter Must customize? Description
service_id Yes (if adding/removing services) Unique integer ID. Must match the corresponding client-side service
max_entries No Log buffer size (uses global anchor)
model_metadata_list No Reference to server_model_list (uses YAML anchor). All servers share the same model list
device Yes PyTorch device string (e.g., cuda:0, cuda:1, cpu). Assign a different GPU to each service for parallelism. Run nvidia-smi to see available GPUs
incoming_zmq_sockname No Bare ZMQ socket name for receiving images from the QUIC server. Auto-resolved to ipc://<zmq_dir>/<name> by server_main.py
incoming_shm_filename No SHM region name for incoming image data
outgoing_zmq_sockname No Bare ZMQ socket name for sending inference results back to the QUIC server
outgoing_shm_filename No SHM region name for outgoing inference results
thread_concurrency No Thread pool size for server operations
shm_filesize No SHM region size in bytes (uses global anchor)
zmq_kill_switch_sockname No Bare ZMQ socket name for graceful shutdown from server_main.py
mock_inference_output_path Maybe Set to a .npz file path (e.g., example_effdet_d4_output.npz) to skip model loading and return pre-recorded detections. Useful for testing without a GPU. Set to null for real inference
mock_model_latency_csv_path Maybe Set to the experiment_model_info.csv path to simulate per-model inference latency in mock mode. Only used when mock_inference_output_path is set. Set to null to skip latency simulation

3. QUIC Configuration — config/quic_config_client.yaml / config/quic_config_gcloud.yaml

Configures the QUIC transport layer (Rust binaries: quic_client and quic_server). You need two copies of this config: one for the client machine and one for the server machine. They share the same schema but may differ in timing parameters and log paths.

Paths and Timing

experiment_output_dir: "/path/to/experiment-out"
zmq_dir: "/path/to/experiment-out/zmq"
quic_client_log_subdir: quic-client-out
quic_server_log_subdir: quic-server-out

slo_timeout_ms: 100
junk_tx_loop_interval_ms: 100
logging_interval_ms: 500
init_allocation: 100000000.0
bw_update_interval_ms: 500
bw_polling_interval_ms: 200
max_junk_payload_Mb: 0.5
junk_restart_interval_ms: 500
Parameter Must customize? Description
experiment_output_dir Yes Base output directory. The QUIC client/server creates a timestamped subdirectory here (e.g., experiment-out/quic_client_2024-01-15_10-30-00/) for each run. Must match the value used in client and server Python configs
zmq_dir Yes Directory for ZMQ IPC socket files. Must match across client, server, and QUIC configs. The QUIC binaries assert this directory exists before starting
quic_client_log_subdir No Subdirectory name within the timestamped run directory for QUIC client Parquet logs. Default quic-client-out
quic_server_log_subdir No Subdirectory name within the timestamped run directory for QUIC server Parquet logs. Default quic-server-out
slo_timeout_ms Maybe Timeout for dropping stale queued frames in the QUIC send loop (ms). Should be aligned with SLO_TIMEOUT in the client config. Frames older than this are silently dropped from the LIFO queue
junk_tx_loop_interval_ms No Interval between junk service transmissions (ms). Lower values probe bandwidth more aggressively. Cloud config uses 7, client config uses 100
logging_interval_ms No How often network statistics are logged to Parquet (ms)
init_allocation No Initial per-service bandwidth allocation in bytes/sec, used before the first LP solver run completes. Default 100000000 (100 MB/s) is intentionally high to avoid dropping early frames
bw_update_interval_ms No How often the QUIC client sends bandwidth/RTT updates to the BandwidthAllocator (ms)
bw_polling_interval_ms No How often the BandwidthAllocator polls for new updates (ms)
max_junk_payload_Mb No Bandwidth limit for the junk service in Mb/s (not payload size per send). Caps the bandwidth consumed by probing traffic
junk_restart_interval_ms No How long the junk service waits before restarting transmission after going idle (ms)

Service List

services: [1, 2, 3, 4]
enable_junk_service: True
Parameter Must customize? Description
services Yes (if adding/removing services) List of all service IDs. If enable_junk_service is True, the highest ID is the junk service and should not appear in the client's service_id_list
enable_junk_service Maybe If True, the last service in the list sends dummy data to probe available bandwidth. Recommended True for accurate bandwidth estimation

Logging Flags (Client-side QUIC)

client_enable_bw_stat_log: True
client_enable_allocation_stat_log: True
client_enable_network_stat_log: True
client_enable_incoming_image_context_log: True
client_enable_outgoing_image_context_log: True

Controls which Parquet log files the QUIC client produces. Set to False to disable specific logs and reduce disk I/O.

Logging Flags (Server-side QUIC)

server_enable_bw_stat_log: True
server_enable_allocation_stat_log: True
server_enable_network_stat_log: True
server_enable_incoming_image_context_log: True
server_enable_outgoing_image_context_log: True

Note: server_enable_bw_stat_log and server_enable_allocation_stat_log exist for structural symmetry but do not produce logs on the server side (bandwidth refresh and allocation only run on the client).

Log Buffer Capacities

image_context_log_capacity: 100
bw_stat_log_capacity: 100
allocation_stat_log_capacity: 100
network_stat_log_capacity: 100

Max records held in memory before flushing to Parquet, per log type. Higher values reduce I/O frequency at the cost of more memory and potential data loss on crash.

Configuration Checklist

When deploying the system, work through this checklist to ensure all paths and parameters are set correctly.

1. Choose an output directory

All components write logs and IPC socket files to a shared output directory tree. Set experiment_output_dir and zmq_dir consistently across all four config files. The orchestrators automatically create timestamped run subdirectories and ZMQ directories at startup — you only need the base directory to exist (or be creatable):

mkdir -p ~/experiment-out

2. Client-side (client_config.yaml)

  • logging_config_filepath points to config/logging_config.yaml (absolute path)
  • experiment_output_dir is set to your chosen output directory
  • zmq_dir is set (e.g., ~/experiment-out/zmq) and matches all other configs
  • DST_IP is set to your cloud server's public IP address
  • parquet_eval_dir points to the directory containing pre-computed utility curve Parquet files
  • model_info_csv_path points to the experiment_model_info.csv file
  • camera_stream_config_list[].usb_id matches your USB camera device IDs (run ls /dev/video*)
  • camera_np_size matches your camera's resolution (default [1080, 1920, 3] for 1080p)
  • Number of entries in main_client_config_list, camera_stream_config_list, and bandwidth_allocator_config.outgoing_zmq_client_socknames all match
  • service_id values are consistent across client, camera, plotter, bandwidth allocator, and QUIC configs
  • If testing without cameras: set mock_camera_image_path to a JPEG file path

3. Server-side (server_config_gcloud.yaml)

  • logging_config_filepath points to config/logging_config.yaml (absolute path)
  • experiment_output_dir is set to your chosen output directory (same as client config)
  • zmq_dir matches all other configs
  • server_model_list[].checkpoint_path points to valid EfficientDet .ckpt files (see MODELS.md)
  • server_config_list[].device assigns a unique GPU to each service (e.g., cuda:0, cuda:1, cuda:2)
  • Number of entries in server_config_list matches the number of real services (not including junk)
  • If testing without a GPU: set mock_inference_output_path to a .npz file path and optionally set mock_model_latency_csv_path to simulate per-model latency

4. QUIC configs (quic_config_client.yaml and quic_config_gcloud.yaml)

  • experiment_output_dir matches the value in client and server configs
  • zmq_dir matches all other configs
  • services list includes all active service IDs plus the junk service (e.g., [1, 2, 3, 4])
  • slo_timeout_ms is aligned with SLO_TIMEOUT in the client config
  • Both client and server QUIC configs have the same services list

5. Cross-config consistency

  • experiment_output_dir is the same across all four config files
  • zmq_dir is the same across all four config files
  • The services list in QUIC configs includes all service_id values from client and server configs, plus the junk service
  • The SLO_TIMEOUT (client) and slo_timeout_ms (QUIC) are aligned
  • SHM filenames in client config match what the QUIC binaries expect (follow the naming convention client-service{N}-incoming-shm / client-service{N}-outgoing-shm)
  • SHM filenames in server config follow the convention server-service{N}-outgoing-shm / server-service{N}-incoming-shm