Merge pull request #4 from intel/iaa_batch_release_v1.0

Updated with changes required for IAA batching benchmarking

Author: bravindr

tests/madvise/README.md

@@ -1,55 +1,51 @@
-# Instructions
-madvise-test suite allows benchmarking of zswap with different compression algorithms. It loads a data corpus (silesia.tar for example), swap-out all the pages, swap-in all the pages in the meanwhile monitoring the swap-in and swap-out latency along with the compressed size using bpftrace.
-
-## Prerequisite
-#### Hardware
-* Intel 4th Gen Intel® Xeon® Scalable Processors or later with IAA enebled. Please see [IAA user guide](https://cdrdv2-public.intel.com/780887/354834_IAA_UserGuide_June23.pdf)for more details about IAA configuration.
-#### Software
-This framework has dependencies on IAA RFC kernel patches. Please see [instructions on building the kernel with right patch sets](https://github.com/intel/memory-usage-analyzer/wiki/Integration-of-IAA-RFC-patches-to-6.12-upstream-kernel))
-
-## Run in Baremetal
-1. Run
-   ```
-    ./make_swap_space.sh
-   ```
-   Additional details of make_swap_space.sh script:
-	Command Line Arguments: The script now accepts -l for specifying the swap file location and -s for specifying the swap size in GB.
-	Default Values: If no arguments are provided, it defaults to /mnt/nvme1/swapfile for the location and 1GB for the size.
-	Dynamic Path Handling: It dynamically checks the available space in the directory derived from the provided or default swap file location.
-	Example of creating 4GB swap space at /mnt/nvme1/swapfile
-	```
-	./make_swap_space.sh  [-l <path_to_swap_file>] [-s <swap_size_in_GBi>]
-	```
-2. Configure IAA device
-    ```
-    ./enable_kernel_iaa
-    ```
-3. Active zswap by runnig
-   ```
-   ./enable_zswap.sh
-   ```
-4. Collect data and generate reports for all the compressors.
-   ```
-   ./collect_all.sh
-   ```
- This will generate html files for CDFs of compress, decompress, compression ratio and page fault latencies and a summary.xlsx file. This will also generate (to stdout) P50 & P99 values.
+# Introducution
+zswap performance benchmarking with IAA uses madvise() system call and MADV_PAGEOUT. It loads the entire dataset (silesia.tar for example) to memory, swap-out all the pages and swap-in all the pages, monitoring the time spent in swap-in and swap-out and other key metrics. There are two benchmarking scenarios
+
+1. benchmark single page
+2. benchmark with IAA batching to take advantage of parallel processing in IAA.
+
+## Prerequisites
+1. Platform with Intel Xeon 4th generation (or higher) processor and IAA.
+2. [Memory Usage Analyzer Framework](https://github.com/intel/memory-usage-analyzer/tree/main?tab=readme-ov-file#install_)
+3. Kernel with IAA RFC patches. Please see instructions on building the kernel [here](https://github.com/intel/memory-usage-analyzer/wiki/Integration-of-IAA-RFC-patches-to-6.12-upstream-kernel).
+
+## Run single-page Microbenchmarks
+
+Collect data and generate reports for all the compressors for single-page. Depending on the number of IAA devices on the system, the setup scripts needs to be modified. The list of compressors and datasets can be modified as needed.
+```
+    # For all 4 devices` per socket
+    ./collect_single_page.sh  | tee single_page.txt
+
+    # For SKUs with only 1 IAA device per socket
+    ./collect_single_page.sh  -d 1  | tee single_page.txt
+
+``` 
+This will generate a summary of the key metrics for each dataset. In addition to that more detailed data points like CDFs and a summary .xls will be generated under the results_* directory
+
+## Run Microbenchmarks with IAA batching
+Collect data and generate reports for all the compressors for batch processing. The list of compressors, datasets and batch sweep can be modifed as needed.
+```
+    # For all 4 devices` per socket
+    ./collect_batch.sh | tee batch.txt
+    # For SKUs with only 1 IAA device per socket
+    ./collect_batch.sh -d 1 | tee batch.txt
+```
+This will generate swap-in and swap-out latency reports for different batches for IAA along with software compressors.
 
 ## Additional details
-For running individual compressors
+For running individual compressors low-level script can be utilized.
 
-   ```
+```
     echo 'lz4' > /sys/module/zswap/parameters/compressor
     ./collect_bpftraces.sh
     echo 'deflate-iaa-canned' > /sys/module/zswap/parameters/compressor
     ./collect_bpftraces.sh
-   ```
+   ``` 
 ... and so on. This will generate a <compressor>_output file for each run
 
 Once all runs are collected, run the post-processing Python script:
-   ```
+```
     ./process_bpftraces.py
-   ```
-This will generate html files for CDFs of compress, decompress, compression ratio and page fault latencies and a summary.xlsx file.
-This will also generate (to stdout) P50 & P99 values.
+```
 
 

tests/madvise/clear_logs.sh

@@ -0,0 +1,8 @@
+#!/bin/bash 
+#SPDX-License-Identifier: BSD-3-Clause
+#Copyright (c) 2025, Intel Corporation
+#Description: clear intermediate logs
+rm -f *_output
+rm -f  *.html
+rm -f *.xlsx
+rm -f perf_*

tests/madvise/collect_bpftraces.sh

@@ -1,25 +1,32 @@
-#!/usr/bin/env bash
+#!/usr/bin/bash
 #SPDX-License-Identifier: BSD-3-Clause
-#Copyright (c) 2023, Intel Corporation
+#Copyright (c) 2025, Intel Corporation
+#Description: Wrapper which runs madvise workload and collect bpftraces
 
 # Define the location of the swap file
 SWAP="silesia.tar"
 COMP=$(cat /sys/module/zswap/parameters/compressor 2>/dev/null)
 CORE_FREQUENCY=2500
+CBATCH=1
+DBATCH=0
+MTHP="4kB"
 
 
-while getopts "f:t:" opt; do
+while getopts "f:t:c:d:m:" opt; do
   case $opt in
     f) CORE_FREQUENCY="$OPTARG" ;;
     t) SWAP="$OPTARG" ;;
+    c) CBATCH="$OPTARG" ;;
+    d) DBATCH="$OPTARG" ;;
+    m) MTHP="$OPTARG" ;;
     \?) echo "Invalid option: -$OPTARG" >&2; exit 1 ;;
   esac
 done
 
 # Function to handle errors
 handle_error() {
     echo "Error: $1"
-    exit 1
+    #exit 1
 }
 
 
@@ -35,7 +42,10 @@ if [ ! -f "$SWAP" ] ; then
         "silesia.tar")
              echo "Downloading $SWAP from http://wanos.co/assets/silesia.tar"
              wget --no-check-certificate http://wanos.co/assets/silesia.tar || handle_error "Failed to download $SWAP" ;;
-          *) echo "Invalid file $SWAP" ;  exit 1 ;
+        "4300_all")
+             echo "using  $SWAP locally" ;;
+          *) echo "Invalid file $SWAP" ;  exit 1 ;;
+	     
      esac
 fi
 
@@ -53,41 +63,92 @@ if [ ! -f madvise_test ]; then
 fi
 
 # Turn off read-ahead to optimize performance
-echo 0 > /proc/sys/vm/page-cluster || handle_error "Failed to set page-cluster"
+#echo 0 > /proc/sys/vm/page-cluster || handle_error "Failed to set page-cluster"
 
 
+if [[ $COMP == 'deflate-iaa-canned' || $COMP == 'deflate-iaa-dynamic'||$COMP == 'deflate-iaa' ]]; then
+    echo "true" > /sys/kernel/mm/swap/singlemapped_ra_enabled
+else
+    echo "false" > /sys/kernel/mm/swap/singlemapped_ra_enabled
+    #CBATCH=1
+    #DBATCH=0
+fi
+
+echo ${DBATCH} > /proc/sys/vm/page-cluster || handle_error "Failed to set page-cluster"
+sysctl vm.compress-batchsize=${CBATCH}  || handle_error "Failed to set compress-batchsize"
+
 # Clear transparent huge pages configuration
 echo 'never' > /sys/kernel/mm/transparent_hugepage/hugepages-2048kB/enabled || handle_error "Failed to clear hugepages-2048kB configuration"
 echo 'never' > /sys/kernel/mm/transparent_hugepage/enabled || handle_error "Failed to clear transparent_hugepage configuration"
 
-# Calculate swap size and pages
-SZ=$(ls -s "$SWAP" | awk '{print $1}')
-NPGS=$(echo "scale=0; $SZ/4-1" | bc -l)
 
 
-CMD="./madvise_test ${SWAP} ${NPGS}"
+# Set mTHP
+mthp_sizes=('16kB' '32kB' '64kB' '128kB' '256kB' '512kB' '1024kB' '2048kB')
+IFS=',' read -a mthp_list <<< "$MTHP"
+#echo ${mthp_list[@]}
+
+for mthp in "${mthp_sizes[@]}"; do
+    echo 'never'> /sys/kernel/mm/transparent_hugepage/hugepages-${mthp}/enabled
+done
+
+max_mthp=0
+for mthp in "${mthp_list[@]}"; do
+    mthp=`echo $mthp | tr -d ' '`
+    if [[ ${mthp_sizes[@]} =~ ($mthp) ]] ; then
+        if [ $mthp != '4kB' ]; then
+            echo "configuring mthp ${mthp}"
+            echo 'always' > /sys/kernel/mm/transparent_hugepage/hugepages-${mthp}/enabled
+	fi
+    fi
+done
+
+
+# Calculate swap size and pages
+SZ=$(ls -s "$SWAP" | awk '{print $1}')
+PAGE_SIZE=${mthp_list[0]}
+PAGE_SIZE=4kB
+echo "page-size:${PAGE_SIZE}"
+PAGE_SIZE=$(echo ${PAGE_SIZE} | awk '{print substr($1,1,length($1)-2)}' )
+NPGS=$(echo "scale=0; $SZ/${PAGE_SIZE}" | bc -l)
+# change it to bytes
+PAGE_SIZE=$(( PAGE_SIZE * 1024))
+
+QAT_ENABLED=`bpftrace -l | grep qat_comp_alg_compress | wc -l`
+
+CMD="./madvise_test ${SWAP} ${NPGS} ${PAGE_SIZE}"
+echo ${CMD}
 if [[ $COMP == 'deflate-iaa-canned' || $COMP == 'deflate-iaa-dynamic'||$COMP == 'deflate-iaa' ]];then
     COMP_STR="iaa_comp_acompress"
     DCOMP_STR="iaa_comp_adecompress"
     SZ_STR="arg0+68"
+    #SZ_STR="((struct acomp_req*)arg0)->dlen"
 elif [ $COMP == 'lzo-rle' ]; then
     COMP_STR="lzorle_scompress"
     DCOMP_STR="lzorle_sdecompress"
     SZ_STR="arg4"
+elif [ $COMP == 'qat_deflate' ] &&  [ ${QAT_ENABLED} -gt 0 ]; then
+    COMP_STR="qat_comp_alg_compress"
+    DCOMP_STR="qat_comp_alg_decompress"
+    SZ_STR="arg4"
 else
     COMP_STR="${COMP}_scompress"
     DCOMP_STR="${COMP}_sdecompress"
     SZ_STR="arg4"
 fi
 
-PROG=`echo "kprobe:${COMP_STR} { @start=nsecs; @sz=${SZ_STR}; } kretprobe:${COMP_STR} { printf (\"C %d\\nR %d\\n\", nsecs-@start, *@sz); }
-            kprobe:${DCOMP_STR} { @start=nsecs; } kretprobe:${DCOMP_STR} { printf (\"D %d\\n\", nsecs-@start); }
-	    kprobe:handle_mm_fault /(arg1&0xfff) == 0/{@pf[cpu]=nsecs; if(arg2!=0x1254) {@pf[cpu]=0;}} kretprobe:handle_mm_fault {if(@pf[cpu]) {printf(\"P %d\\n\", nsecs-@pf[cpu]);} @pf[cpu]=0;}
-	    kprobe:swap_writepage { @start_swap_write=nsecs; } kretprobe:swap_writepage { printf (\"SW %d\\n\", nsecs-@start_swap_write); }
-	    kprobe:swap_read_folio { @start_swap_read=nsecs; } kretprobe:swap_read_folio { printf (\"SR %d\\n\", nsecs-@start_swap_read);}
-	    kprobe:zswap_compress { @start_zswap_comp=nsecs; @zsz=arg1;} kretprobe:zswap_compress { printf (\"ZC %d\\nZCSZ %d\\nZCR %d\\n\", nsecs-@start_zswap_comp, *(@zsz+8), retval&0x1);}
-	    kprobe:zswap_decompress { @start_zswap_decomp=nsecs; } kretprobe:zswap_decompress { printf (\"ZD %d\\n\", nsecs-@start_zswap_decomp);}
-	    END { clear(@pf); delete(@start); delete(@sz);}"`
+
+# TODO:  count errors from swap_crypto_acomp_compress_batch
+
+PROG=`echo "kprobe:${COMP_STR} { @start=nsecs; @sz=${SZ_STR}; } kretprobe:${COMP_STR} { printf (\"C %d\\nR %d\\n\", nsecs-@start, *@sz); } 
+            kprobe:${DCOMP_STR} { @start=nsecs; } kretprobe:${DCOMP_STR} { printf (\"D %d\\n\", nsecs-@start); } 
+
+	    kprobe:swap_writepage { @start_swap_write=nsecs; } kretprobe:swap_writepage { printf (\"SW %d\\n\", nsecs-@start_swap_write); } 
+	    kprobe:swap_read_folio { @start_swap_read=nsecs; } kretprobe:swap_read_folio { printf (\"SR %d\\n\", nsecs-@start_swap_read);} 
+
+	    END { delete(@start); delete(@sz);}"`
 
 perf stat -o perf_${COMP}.log bpftrace -e "${PROG}" -c "${CMD}" -o ${COMP}_output
 echo "Script completed successfully."
+
+

tests/madvise/disable_iaa

@@ -1,6 +1,8 @@
 #!/usr/bin/env bash
 #SPDX-License-Identifier: BSD-3-Clause
-#Copyright (c) 2023, Intel Corporation
+#Copyright (c) 2025, Intel Corporation
+#Description: Disable all IAA devices
+
 
 IAX_CONFIG_PATH=/sys/bus/dsa/devices
 
@@ -10,6 +12,7 @@ IAX_BIND_WQ_PATH=/sys/bus/dsa/drivers/crypto/bind
 IAX_UNBIND_PATH=/sys/bus/dsa/drivers/idxd/unbind
 IAX_UNBIND_WQ_PATH=/sys/bus/dsa/drivers/crypto/unbind
 
+echo lzo > /sys/module/zswap/parameters/compressor
 #
 # count iax instances
 #
@@ -34,3 +37,9 @@ done
 
 rmmod iaa_crypto
 modprobe iaa_crypto
+
+
+
+
+
+

tests/madvise/enable_kernel_iaa

@@ -1,6 +1,7 @@
-#!/usr/bin/env bash
+#!/bin/bash
 #SPDX-License-Identifier: BSD-3-Clause
-#Copyright (c) 2023, Intel Corporation
+#Copyright (c) 2025, Intel Corporation
+#Description: Configure zswap
 
 # Enable zswap
 echo "Enabling zswap..."