Update v5e perf results and fix rto_setup for GCE

README.md

@@ -233,15 +233,24 @@ For a 52B model. See full run configs in `MaxText/configs/` as `1xv4-384.sh` and
 
 For a 16B model. See full run configs in `MaxText/configs/` as `16b.sh`.
 
-| Hardware | TFLOP/sec/chip | MFU |
-| ----------- | ---------------- | ----- |
-| 2x v5e-256 | 114 | 57.8% |
+| Hardware | TFLOP/sec/chip | MFU |
+| ----------- | ---------------- | ------ |
+| 1x v5e-256 | 120 | 67.10% |
+| 2x v5e-256 | 117 | 59.37% |
+| 4x v5e-256 | 117 | 59.14% |
+| 8x v5e-256 | 115 | 58.27% |
+| 16x v5e-256 | 111 | 56.56% |
 
 For a 32B model. See full run configs in `MaxText/configs/` as `32b.sh`.
 
-| Hardware | TFLOP/sec/chip | MFU |
-| ----------- | ---------------- | ----- |
-| 2x v5e-256 | 113 | 57.3% |
+| Hardware | TFLOP/sec/chip | MFU |
+| ----------- | ---------------- | ------ |
+| 1x v5e-256 | 132 | 66.86% |
+| 2x v5e-256 | 128 | 64.81% |
+| 4x v5e-256 | 126 | 64.10% |
+| 8x v5e-256 | 125 | 63.67% |
+| 16x v5e-256 | 123 | 62.26% |
+
 
 More details on reproducing these 16B and 32B model results on v5e can be found in `v5e_16b_32b.md`.
 

rto_setup.sh

@@ -1,9 +1,15 @@
-echo "Running rto_setup.sh..."
+echo "Running rto_setup.sh for GCE..."
 
 # Stop execution if any command exits with error
 set -e
 
+echo "Adjust RTO and apply non cache copy"
 first_line_res=$(ip route show | head -n 1)
-sudo ip route change ${first_line_res} rto_min 5ms
-sudo ethtool -K ens9 tx-nocache-copy on
-echo "rto_setup finished"
+if [[ "$(echo "$first_line_res" | grep "rto_min lock 5ms" | wc -l)" -eq 0 ]]; then
+ sudo ip route change "${first_line_res}" rto_min 5ms
+fi
+dev_name=$(echo "$first_line_res" | awk -F'[[:space:]]' '{ print $5 }')
+echo "dev_name=${dev_name}"
+sudo ethtool -K "${dev_name}" tx-nocache-copy on
+
+echo "rto_setup.sh finished"