carm-paraver 1.0.0.dev0__tar.gz → 1.0.0.dev1__tar.gz

carm_paraver-1.0.0.dev1/PKG-INFO

@@ -0,0 +1,168 @@
+Metadata-Version: 2.4
+Name: carm-paraver
+Version: 1.0.0.dev1
+Summary: Dash-based CARM analysis for Paraver traces
+Author: CARM Contributors
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: dash>=4.1.0
+Requires-Dist: dash-bootstrap-components>=2.0.4
+Requires-Dist: dash-daq>=0.6.0
+Requires-Dist: numpy>=2.0.0
+Requires-Dist: pandas>=2.3.3
+Requires-Dist: platformdirs>=4.2.2
+Requires-Dist: plotly>=6.0.0
+Dynamic: license-file
+
+# CARM-Paraver GUI
+
+This GUI allows the analysis of [Paraver](https://tools.bsc.es/paraver) traces on the Cache-Aware Roofline Model (CARM) for floating-point operations. It can be launched from the Paraver interface and send labeled events back to Paraver for visualization.
+
+# Requirements
+- Python (tested with 3.9.25, 3.10.12, 3.12.3)
+- [Paraver, Extrae](https://tools.bsc.es/downloads)
+
+# How to use
+
+## Installation
+**The recommended way to install the package** is via `pip`:
+```bash
+pip install carm-paraver
+```
+Alternatively, you can install it from source by cloning this repository and running:
+```bash
+pip install .
+```
+If the install fails due to dependency conflicts, you can use a Python virtual environment to install the package and its dependencies in an isolated environment. To do this, you can run:
+```bash
+python -m venv .venv
+source .venv/bin/activate
+pip install carm-paraver
+```
+If you install in a virtual environment, make sure to run Paraver from the same environment:
+```bash
+source .venv/bin/activate
+wxparaver
+```
+
+## First-time Setup
+CARM-Paraver needs `paramedir` to be in your PATH in order to run. To add it, add Paraver's bin directory to your PATH. You can make this permanent by appending it to your `.bashrc` or `.bash_profile` (change the path accordingly):
+
+```bash
+export PATH=/path/to/paraver/bin:$PATH
+```
+
+## Running
+The GUI is launched via the Paraver interface like so:
+1. Use [Extrae](https://github.com/bsc-performance-tools/extrae) to generate a trace with the required counters ([see how to configure Extrae below](#paraver-trace-requirements)).
+2. Load the trace in Paraver, and zoom into a section of interest.
+3. Right click the timeline and select the option to launch the CARM GUI.
+4. Configure the options in Paraver to your liking (see [Launch Configuration](#launch-configuration)), and click "Run".
+5. Click the link printed in the Paraver console to open the GUI in your browser.
+
+You will now have the CARM GUI open, showing the architecture's roofline, and the events from the Paraver trace represented as points on the plot. Their position on the roofline, which is determined by their performance and arithmetic intensity, can be used to identify bottlenecks and optimization opportunities for the respective code section. Check the [CARM GUI Features](#carm-gui-features) section for more details about the GUI, and how you can label events and send them back to Paraver for visualization.
+
+If you get any errors, be sure to consult the [First-time Setup](#first-time-setup) and [Paraver Trace Requirements](#paraver-trace-requirements) sections.
+
+## Paraver Trace Requirements
+
+To enable CARM analysis, your Paraver trace needs to include information on the floating-point and memory operations performed by the application. To do this, [configure Extrae](https://tools.bsc.es/doc/html/extrae/xml.html#xml-section-performance-counters) to include the counters in the tables below.
+
+#### Which counters to include?
+Include only the necessary counters for your analysis, so they fit in a single counter set. If too many counters are active, accuracy may be reduced.
+
+Take the application examples below. For each case, the tables below indicate which counters you should include in your Extrae configuration:
+- **App 1**: The application only uses double precision, but you don't know which vector ISAs it uses.
+- **App 2**: The application is vectorized with AVX2, using both precisions.
+
+If you are unsure, include all counters and prune them later as you learn more about the application. Using separate load and store counters is recommended, as it allows for a more detailed analysis.
+
+#### Intel CPUs
+| FP/Mem Operation | Intel Counter                              | App 1   | App 2   |
+| ---------------- | ------------------------------------------ | ------- | ------- |
+| Scalar DP Insts  | `FP_ARITH_INST_RETIRED:SCALAR_DOUBLE`      | ✓ | ✓ |
+| Scalar SP Insts  | `FP_ARITH_INST_RETIRED:SCALAR_SINGLE`      |         | ✓ |
+| SSE DP Insts     | `FP_ARITH_INST_RETIRED:128B_PACKED_DOUBLE` | ✓ |         |
+| SSE SP Insts     | `FP_ARITH_INST_RETIRED:128B_PACKED_SINGLE` |         |         |
+| AVX2 DP Insts    | `FP_ARITH_INST_RETIRED:256B_PACKED_DOUBLE` | ✓ | ✓ |
+| AVX2 SP Insts    | `FP_ARITH_INST_RETIRED:256B_PACKED_SINGLE` |         | ✓ |
+| AVX512 DP Insts  | `FP_ARITH_INST_RETIRED:512B_PACKED_DOUBLE` | ✓ |         |
+| AVX512 SP Insts  | `FP_ARITH_INST_RETIRED:512B_PACKED_SINGLE` |         |         |
+| Loads            | `MEM_INST_RETIRED:ALL_LOADS`               | ✓ | ✓ |
+| Stores           | `MEM_INST_RETIRED:ALL_STORES`              | ✓ | ✓ |
+| Loads and Stores | `MEM_INST_RETIRED:ALL`                     |         |         |
+
+#### AMD CPUs
+| FP/Mem Operation | AMD Counter                                    | App 1   | App 2   |
+| ---------------- | ---------------------------------------------- | ------- | ------- |
+| Mul/Add DP Flops | `retired_sse_avx_operations:dp_mult_add_flops` | ✓ | ✓ |
+| Mul/Add SP Flops | `retired_sse_avx_operations:sp_mult_add_flops` |         | ✓ |
+| Add/Sub DP Flops | `retired_sse_avx_operations:dp_add_sub_flops`  | ✓ | ✓ |
+| Add/Sub SP Flops | `retired_sse_avx_operations:sp_add_sub_flops`  |         | ✓ |
+| Mul DP Flops     | `retired_sse_avx_operations:dp_mult_flops`     | ✓ | ✓ |
+| Mul SP Flops     | `retired_sse_avx_operations:sp_mult_flops`     |         | ✓ |
+| Div DP Flops     | `retired_sse_avx_operations:dp_div_flops`      | ✓ | ✓ |
+| Div SP Flops     | `retired_sse_avx_operations:sp_div_flops`      |         | ✓ |
+| Loads            | `ls_dispatch:ld_dispatch`                      | ✓ | ✓ |
+| Stores           | `ls_dispatch:store_dispatch`                   | ✓ | ✓ |
+
+#### Additional recommendations
+For best results, when labeling your code with [Extrae events](https://tools.bsc.es/doc/html/extrae/api.html), e.g. with `Extrae_eventandcounters` calls, **avoid labeling regions that include MPI calls**. Focus on labeling regions of pure computation, as MPI calls will cause the region and hardware counter timestamps to not match, preventing them from being shown on the CARM GUI.
+
+## CARM Benchmarking
+
+To benchmark your architecture and display its roofline in the CARM GUI, use the [CARM Tool](https://github.com/champ-hub/carm-roofline). **Note: for compatibility, use the [latest version of the CARM Tool](https://pypi.org/project/carm-roofline/)**
+
+This tool ships a series of sample rooflines from a MareNostrum 5 GPP node.
+
+## CARM GUI Features
+
+### Launch Configuration
+**Use window colors:**
+Controls which coloring scheme is used in the CARM GUI: the same colors as the Paraver timeline (if enabled) or the selected CARM GUI coloring scheme (see right sidebar options).
+
+**Use Semantic Window:**
+Controls whether the Paraver semantic window is used: if enabled, the GUI displays only the timestamps that are within the semantic window of the Paraver timeline. If disabled, all timestamps in the trace are displayed.
+
+**Accumulate values:**
+Controls whether timestamps (with the same underlying Paraver value) are averaged. Allows for similar timestamps to be grouped into a single, per-thread point, or to plot all timestamps individually.
+
+### Left Sidebar
+
+**Use Paraver/CARM Colors:**
+Same as above's "Use window colors"
+
+**Use Semantic Window / All Timestamps:**
+Same as above's "Use Semantic Window"
+
+**Plot Raw/Accumulated Values:**
+Same as above's "Accumulate values"
+
+**Re-Sync Timeline With Paraver:**
+Re-syncs the plotted timestamps in the CARM GUI with the timestamps being viewed in the Paraver timeline from which the CARM GUI was launched. This first requires the **Time Sync** button to be clicked on the Paraver side, the CARM GUI will usually keep itself synced to the Paraver timeline whenever the **Time Sync** button is clicked in the Paraver interface. In case the user changes the displayed timestamps in the CARM GUI and wishes to return to the same interval that they have in the Paraver timeline, they can use the **Re-Sync Timeline With Paraver** button.
+
+**Send Timestamps Roof Labels:**
+Labels the timestamps based on which roof they are under, for viewing in Paraver. The path of the generated trace will be printed in the Paraver console, and can be clicked to open the trace in Paraver. You can then select the trace and click *New single timeline window* to view the timestamps with the new labels.
+
+**Send Timestamps LD/ST Percentage Colors:**
+Same as above, but labels the timestamps based on the percentage of loads to stores.
+
+**Send Timestamps SP/DP Percentage Colors:**
+Same as above, but labels the timestamps based on the percentage of single to double precision operations.
+
+### Right Sidebar
+The right sidebar controls the CARM GUI specific features, which include various filtering and coloring options as well as graphical annotations.
+
+Useful options include:
+- **Filter points** by vector ISA or precision
+- **Color points** based on thread ID, precision, vector ISA or load/store ratio
+    - Note that this requires the left sidebar option to be set to "Use CARM GUI Colors".
+
+The plot can be configured to normalize the performance roof to the number of threads. The normalized roofs represent the performance per thread, which matches the Paraver timestamps (also per thread). This mode is recommended when relating application performance to the underlying hardware. The non-normalized roofs represent the overall performance of the architecture, and is best for understanding the hardware capabilities.
+
+## GUI Performance
+The GUI may become slow when plotting a very large number of events. To improve performance, you can:
+- Enable the "Accumulate values" option to group similar events into a single point.
+- Enable the "Use Semantic Window" option to only plot events visible in Paraver.
+- Focus your analysis on a smaller time window in the Paraver timeline.

carm_paraver-1.0.0.dev1/README.md

@@ -0,0 +1,151 @@
+# CARM-Paraver GUI
+
+This GUI allows the analysis of [Paraver](https://tools.bsc.es/paraver) traces on the Cache-Aware Roofline Model (CARM) for floating-point operations. It can be launched from the Paraver interface and send labeled events back to Paraver for visualization.
+
+# Requirements
+- Python (tested with 3.9.25, 3.10.12, 3.12.3)
+- [Paraver, Extrae](https://tools.bsc.es/downloads)
+
+# How to use
+
+## Installation
+**The recommended way to install the package** is via `pip`:
+```bash
+pip install carm-paraver
+```
+Alternatively, you can install it from source by cloning this repository and running:
+```bash
+pip install .
+```
+If the install fails due to dependency conflicts, you can use a Python virtual environment to install the package and its dependencies in an isolated environment. To do this, you can run:
+```bash
+python -m venv .venv
+source .venv/bin/activate
+pip install carm-paraver
+```
+If you install in a virtual environment, make sure to run Paraver from the same environment:
+```bash
+source .venv/bin/activate
+wxparaver
+```
+
+## First-time Setup
+CARM-Paraver needs `paramedir` to be in your PATH in order to run. To add it, add Paraver's bin directory to your PATH. You can make this permanent by appending it to your `.bashrc` or `.bash_profile` (change the path accordingly):
+
+```bash
+export PATH=/path/to/paraver/bin:$PATH
+```
+
+## Running
+The GUI is launched via the Paraver interface like so:
+1. Use [Extrae](https://github.com/bsc-performance-tools/extrae) to generate a trace with the required counters ([see how to configure Extrae below](#paraver-trace-requirements)).
+2. Load the trace in Paraver, and zoom into a section of interest.
+3. Right click the timeline and select the option to launch the CARM GUI.
+4. Configure the options in Paraver to your liking (see [Launch Configuration](#launch-configuration)), and click "Run".
+5. Click the link printed in the Paraver console to open the GUI in your browser.
+
+You will now have the CARM GUI open, showing the architecture's roofline, and the events from the Paraver trace represented as points on the plot. Their position on the roofline, which is determined by their performance and arithmetic intensity, can be used to identify bottlenecks and optimization opportunities for the respective code section. Check the [CARM GUI Features](#carm-gui-features) section for more details about the GUI, and how you can label events and send them back to Paraver for visualization.
+
+If you get any errors, be sure to consult the [First-time Setup](#first-time-setup) and [Paraver Trace Requirements](#paraver-trace-requirements) sections.
+
+## Paraver Trace Requirements
+
+To enable CARM analysis, your Paraver trace needs to include information on the floating-point and memory operations performed by the application. To do this, [configure Extrae](https://tools.bsc.es/doc/html/extrae/xml.html#xml-section-performance-counters) to include the counters in the tables below.
+
+#### Which counters to include?
+Include only the necessary counters for your analysis, so they fit in a single counter set. If too many counters are active, accuracy may be reduced.
+
+Take the application examples below. For each case, the tables below indicate which counters you should include in your Extrae configuration:
+- **App 1**: The application only uses double precision, but you don't know which vector ISAs it uses.
+- **App 2**: The application is vectorized with AVX2, using both precisions.
+
+If you are unsure, include all counters and prune them later as you learn more about the application. Using separate load and store counters is recommended, as it allows for a more detailed analysis.
+
+#### Intel CPUs
+| FP/Mem Operation | Intel Counter                              | App 1   | App 2   |
+| ---------------- | ------------------------------------------ | ------- | ------- |
+| Scalar DP Insts  | `FP_ARITH_INST_RETIRED:SCALAR_DOUBLE`      | ✓ | ✓ |
+| Scalar SP Insts  | `FP_ARITH_INST_RETIRED:SCALAR_SINGLE`      |         | ✓ |
+| SSE DP Insts     | `FP_ARITH_INST_RETIRED:128B_PACKED_DOUBLE` | ✓ |         |
+| SSE SP Insts     | `FP_ARITH_INST_RETIRED:128B_PACKED_SINGLE` |         |         |
+| AVX2 DP Insts    | `FP_ARITH_INST_RETIRED:256B_PACKED_DOUBLE` | ✓ | ✓ |
+| AVX2 SP Insts    | `FP_ARITH_INST_RETIRED:256B_PACKED_SINGLE` |         | ✓ |
+| AVX512 DP Insts  | `FP_ARITH_INST_RETIRED:512B_PACKED_DOUBLE` | ✓ |         |
+| AVX512 SP Insts  | `FP_ARITH_INST_RETIRED:512B_PACKED_SINGLE` |         |         |
+| Loads            | `MEM_INST_RETIRED:ALL_LOADS`               | ✓ | ✓ |
+| Stores           | `MEM_INST_RETIRED:ALL_STORES`              | ✓ | ✓ |
+| Loads and Stores | `MEM_INST_RETIRED:ALL`                     |         |         |
+
+#### AMD CPUs
+| FP/Mem Operation | AMD Counter                                    | App 1   | App 2   |
+| ---------------- | ---------------------------------------------- | ------- | ------- |
+| Mul/Add DP Flops | `retired_sse_avx_operations:dp_mult_add_flops` | ✓ | ✓ |
+| Mul/Add SP Flops | `retired_sse_avx_operations:sp_mult_add_flops` |         | ✓ |
+| Add/Sub DP Flops | `retired_sse_avx_operations:dp_add_sub_flops`  | ✓ | ✓ |
+| Add/Sub SP Flops | `retired_sse_avx_operations:sp_add_sub_flops`  |         | ✓ |
+| Mul DP Flops     | `retired_sse_avx_operations:dp_mult_flops`     | ✓ | ✓ |
+| Mul SP Flops     | `retired_sse_avx_operations:sp_mult_flops`     |         | ✓ |
+| Div DP Flops     | `retired_sse_avx_operations:dp_div_flops`      | ✓ | ✓ |
+| Div SP Flops     | `retired_sse_avx_operations:sp_div_flops`      |         | ✓ |
+| Loads            | `ls_dispatch:ld_dispatch`                      | ✓ | ✓ |
+| Stores           | `ls_dispatch:store_dispatch`                   | ✓ | ✓ |
+
+#### Additional recommendations
+For best results, when labeling your code with [Extrae events](https://tools.bsc.es/doc/html/extrae/api.html), e.g. with `Extrae_eventandcounters` calls, **avoid labeling regions that include MPI calls**. Focus on labeling regions of pure computation, as MPI calls will cause the region and hardware counter timestamps to not match, preventing them from being shown on the CARM GUI.
+
+## CARM Benchmarking
+
+To benchmark your architecture and display its roofline in the CARM GUI, use the [CARM Tool](https://github.com/champ-hub/carm-roofline). **Note: for compatibility, use the [latest version of the CARM Tool](https://pypi.org/project/carm-roofline/)**
+
+This tool ships a series of sample rooflines from a MareNostrum 5 GPP node.
+
+## CARM GUI Features
+
+### Launch Configuration
+**Use window colors:**
+Controls which coloring scheme is used in the CARM GUI: the same colors as the Paraver timeline (if enabled) or the selected CARM GUI coloring scheme (see right sidebar options).
+
+**Use Semantic Window:**
+Controls whether the Paraver semantic window is used: if enabled, the GUI displays only the timestamps that are within the semantic window of the Paraver timeline. If disabled, all timestamps in the trace are displayed.
+
+**Accumulate values:**
+Controls whether timestamps (with the same underlying Paraver value) are averaged. Allows for similar timestamps to be grouped into a single, per-thread point, or to plot all timestamps individually.
+
+### Left Sidebar
+
+**Use Paraver/CARM Colors:**
+Same as above's "Use window colors"
+
+**Use Semantic Window / All Timestamps:**
+Same as above's "Use Semantic Window"
+
+**Plot Raw/Accumulated Values:**
+Same as above's "Accumulate values"
+
+**Re-Sync Timeline With Paraver:**
+Re-syncs the plotted timestamps in the CARM GUI with the timestamps being viewed in the Paraver timeline from which the CARM GUI was launched. This first requires the **Time Sync** button to be clicked on the Paraver side, the CARM GUI will usually keep itself synced to the Paraver timeline whenever the **Time Sync** button is clicked in the Paraver interface. In case the user changes the displayed timestamps in the CARM GUI and wishes to return to the same interval that they have in the Paraver timeline, they can use the **Re-Sync Timeline With Paraver** button.
+
+**Send Timestamps Roof Labels:**
+Labels the timestamps based on which roof they are under, for viewing in Paraver. The path of the generated trace will be printed in the Paraver console, and can be clicked to open the trace in Paraver. You can then select the trace and click *New single timeline window* to view the timestamps with the new labels.
+
+**Send Timestamps LD/ST Percentage Colors:**
+Same as above, but labels the timestamps based on the percentage of loads to stores.
+
+**Send Timestamps SP/DP Percentage Colors:**
+Same as above, but labels the timestamps based on the percentage of single to double precision operations.
+
+### Right Sidebar
+The right sidebar controls the CARM GUI specific features, which include various filtering and coloring options as well as graphical annotations.
+
+Useful options include:
+- **Filter points** by vector ISA or precision
+- **Color points** based on thread ID, precision, vector ISA or load/store ratio
+    - Note that this requires the left sidebar option to be set to "Use CARM GUI Colors".
+
+The plot can be configured to normalize the performance roof to the number of threads. The normalized roofs represent the performance per thread, which matches the Paraver timestamps (also per thread). This mode is recommended when relating application performance to the underlying hardware. The non-normalized roofs represent the overall performance of the architecture, and is best for understanding the hardware capabilities.
+
+## GUI Performance
+The GUI may become slow when plotting a very large number of events. To improve performance, you can:
+- Enable the "Accumulate values" option to group similar events into a single point.
+- Enable the "Use Semantic Window" option to only plot events visible in Paraver.
+- Focus your analysis on a smaller time window in the Paraver timeline.

{carm_paraver-1.0.0.dev0 → carm_paraver-1.0.0.dev1}/carm_paraver/GUI_utils.py

@@ -362,6 +362,18 @@ def calculate_roofline(values, min_ai):
     FPaidots = [0] * 2
     FPgflopdots = [0] * 2
 
+    try:
+        fp_fma = float(values[5])
+    except (TypeError, ValueError):
+        fp_fma = 0.0
+    try:
+        fp_base = float(values[4])
+    except (TypeError, ValueError):
+        fp_base = 0.0
+
+    # Fall back to non-FMA peak when FP_FMA is missing/zero.
+    fp_peak = fp_fma if fp_fma > 0 else fp_base
+
     ai = np.linspace(min(0.00390625, min_ai), 256, num=200000)
     cache_levels = ["L1", "L2", "L3", "DRAM"]
 
@@ -371,7 +383,7 @@ def calculate_roofline(values, min_ai):
         if values[cache_levels.index(cache_level)] > 0:
             aidots = [0, 0, 0]
             # Compute the first point
-            y_values = carm_eq(ai, values[cache_levels.index(cache_level)], values[5])
+            y_values = carm_eq(ai, values[cache_levels.index(cache_level)], fp_peak)
 
             # Find the point where y_values stops increasing or reaches a plateau
             for i in range(1, len(y_values)):
@@ -531,7 +543,17 @@ def draw_annotation(
 
     if cache_level in cache_levels and values[cache_levels.index(cache_level)] > 0:
         aidots[0] = 0.00390625
-        y_values = carm_eq(ai, values[cache_levels.index(cache_level)], values[5])
+        try:
+            fp_fma = float(values[5])
+        except (TypeError, ValueError):
+            fp_fma = 0.0
+        try:
+            fp_base = float(values[4])
+        except (TypeError, ValueError):
+            fp_base = 0.0
+        fp_peak = fp_fma if fp_fma > 0 else fp_base
+
+        y_values = carm_eq(ai, values[cache_levels.index(cache_level)], fp_peak)
         gflopdots[0] = y_values[0]
         for i in range(1, len(y_values)):
             if y_values[i - 1] == y_values[i]:

{carm_paraver-1.0.0.dev0 → carm_paraver-1.0.0.dev1}/carm_paraver/Paraver_CARM.py

@@ -10,12 +10,14 @@ import logging
 import math
 import os
 import re
+import shutil
 import signal
 import socket
 import subprocess
 import sys
 import tempfile
 import time
+from importlib import resources
 from typing import Any
 
 import dash
@@ -26,6 +28,7 @@ import dash_daq as daq
 # Run: pip install dash dash-bootstrap-components dash-daq numpy pandas plotly
 # To get all of the Libraries in case requirements.txt method fails
 import pandas as pd
+import platformdirs
 import plotly.graph_objects as go
 from dash import ALL, Input, Output, State, callback_context, dcc, html
 from dash.exceptions import PreventUpdate
@@ -106,7 +109,45 @@ if SELECTED_PORT is None:
 
 script_dir = os.path.dirname(os.path.abspath(__file__))
 assets_dir = os.path.join(script_dir, "assets")
-carm_results_path = os.path.join(script_dir, "carm_results", "roofline")
+
+
+def _resolve_roofline_data_dir() -> str:
+    data_dir = platformdirs.user_data_dir("carm", appauthor=False)
+    roofline_dir = os.path.join(data_dir, "roofline")
+    os.makedirs(roofline_dir, exist_ok=True)
+    return roofline_dir
+
+
+def _seed_roofline_data(roofline_dir: str) -> None:
+    if any(name.endswith(".csv") for name in os.listdir(roofline_dir)):
+        return
+
+    sample_ref = resources.files("carm_paraver").joinpath(
+        "sample_data",
+        "roofline",
+        "MN5_roofline.csv",
+    )
+    try:
+        with resources.as_file(sample_ref) as sample_path:
+            shutil.copy2(sample_path, os.path.join(roofline_dir, sample_path.name))
+    except FileNotFoundError:
+        print(
+            "ERROR: bundled MN5 roofline sample is missing; unable to seed data directory.",
+            file=sys.stderr,
+            flush=True,
+        )
+        sys.exit(1)
+    except OSError as exc:
+        print(
+            f"ERROR: unable to seed roofline data in {roofline_dir}: {exc}",
+            file=sys.stderr,
+            flush=True,
+        )
+        sys.exit(1)
+
+
+carm_results_path = _resolve_roofline_data_dir()
+_seed_roofline_data(carm_results_path)
 
 # Global Variables
 n_segments = 0
@@ -278,9 +319,15 @@ parser.add_argument("--mask_csv", action="store_true", help="Use mask CSV")
 parser.add_argument("-ac", action="store_true", help="Optional flag for accumulate values mode")
 parser.add_argument("--csv", type=str, required=True, help="Path to the mask CSV")
 parser.add_argument("trace_path", type=str, help="Path to the .prv file")
+parser.add_argument("--debug", "-d", action="store_true", help="Enable debug logging")
 
 args = parser.parse_args()
 
+if args.debug:
+    logging.basicConfig(level=logging.DEBUG)
+
+logging.debug(f"Parsed arguments: {args}")
+
 min_dur = args.min_dur
 use_paraver_coloring = args.color_csv
 use_mask_csv = args.mask_csv
@@ -434,10 +481,13 @@ if prv_trace_path.endswith(".prv") or prv_trace_path.endswith(".gz"):
     print("Paramedir execution finished, calculating CARM metrics.", flush=True)
 
 # Get CARM results
-if os.path.exists(carm_results_path):
-    csv_files = [f for f in os.listdir(carm_results_path) if f.endswith("_roofline.csv")]
-else:
-    print("ERROR: No CARM results found. Please add them to the ./carm-results/roofline folder.")
+csv_files = sorted(f for f in os.listdir(carm_results_path) if f.endswith("_roofline.csv"))
+if not csv_files:
+    print(
+        f"ERROR: No CARM roofline results found in {carm_results_path}. Add files named *_roofline.csv.",
+        file=sys.stderr,
+        flush=True,
+    )
     sys.exit(1)
 
 # Extract machine names from filenames
@@ -728,6 +778,7 @@ for row in counter_data_df.itertuples(index=False):
     processed += 1
     duration = row.Duration * scaling_unit
     timestamp = row.Timestamp
+    # if FLOP counters are all zero or NaN, skip calculations and set metrics to zero/defaults
     if all(pd.isnull(getattr(row, col)) or getattr(row, col) == 0 for col in columns_to_check):
         no_flops += 1
         full_base_statistics["ThreadID"].append(row.ThreadID)
@@ -2035,12 +2086,16 @@ def update_slider_from_csv(
     current_values,
     selected_file,
 ):
+    def prevent_update_for_reason(reason: str):
+        logging.debug(f"Preventing update on update_slider_from_csv: {reason}")
+        raise PreventUpdate
+
     global sync_csv_path
     global current_file_timestamps
     if mask_button_offset == -1:
-        raise PreventUpdate
+        prevent_update_for_reason("Mask button offset is -1.")
     if not selected_file:
-        raise PreventUpdate
+        prevent_update_for_reason("No file selected.")
     else:
         global no_sync
         global first_load
@@ -2049,7 +2104,7 @@ def update_slider_from_csv(
             new_timestamps = [float(csv_df.iloc[0, 0]), float(csv_df.iloc[1, 0])]
         except Exception:
             first_load += 1
-            raise PreventUpdate from None
+            new_timestamps = current_file_timestamps
 
         ctx = callback_context
         if not ctx.triggered:
@@ -2057,13 +2112,13 @@ def update_slider_from_csv(
         trigger_id = ctx.triggered[0]["prop_id"].split(".")[0]
 
         if new_timestamps == current_file_timestamps and trigger_id != "button-paraver-sync":
-            raise PreventUpdate
+            prevent_update_for_reason("Timestamps in CSV have not changed and trigger is not sync button.")
 
         first_load += 1
         current_file_timestamps = new_timestamps
 
         if first_load <= 1:
-            raise PreventUpdate
+            prevent_update_for_reason("First load.")
 
         try:
             start_index = (full_base_statistics_df["Timestamp"] - new_timestamps[0]).abs().idxmin()
@@ -2114,8 +2169,11 @@ def update_slider_from_csv(
 
         new_slider_indices = [int(new_start_index), int(new_end_index)]
 
+        def print_separator():
+            print("-" * 50, flush=True)
+
         if trigger_id == "button-paraver-sync":
-            print("----------------------------------------------", flush=True)
+            print_separator()
             print(
                 "Sync Button Clicked, updating slider to timestamp range {} - {}".format(
                     filtered_base.loc[new_start_index, "Timestamp"],
@@ -2141,12 +2199,12 @@ def update_slider_from_csv(
                     flush=True,
                 )
 
-            print("----------------------------------------------", flush=True)
+            print_separator()
             no_sync = True
             return new_slider_indices, new_slider_indices, new_timestamps
 
         if new_slider_indices != current_values:
-            print("----------------------------------------------", flush=True)
+            print_separator()
             print(
                 "Sync CSV values changed, updating slider to timestamp range {} - {}".format(
                     filtered_base.loc[new_start_index, "Timestamp"],
@@ -2172,7 +2230,7 @@ def update_slider_from_csv(
                     flush=True,
                 )
 
-            print("----------------------------------------------", flush=True)
+            print_separator()
             no_sync = True
             return new_slider_indices, new_slider_indices, new_timestamps
 
@@ -3946,12 +4004,11 @@ def update_slider_marks(
     triggered_id = ctx.triggered[0]["prop_id"].split(".")[0]
     reset_view = current_values is None or triggered_id in SLIDER_MARKS_CONFIG["value"]["reset_triggers"]
 
-    grouped_count = len(_group_slider_segments(selected_segments, group_value)) if selected_segments else 0
-    max_index = max(grouped_count - 1, 0)
-    if grouped_count < max_dots_auto:
-        initial_range = [0, max_index] if max_index > 0 else [0, 0]
+    if selected_segments:
+        grouped_count = len(_group_slider_segments(selected_segments, group_value))
+        initial_range = [0, max(grouped_count - 1, 0)]
     else:
-        initial_range = [0, min(max_index, 1)] if max_index > 0 else [0, 0]
+        initial_range = [0, 0]
 
     return _resolve_slider_marks_result(
         selected_segments,