Turn notes on Koji hardware into slideware

[riscv-talks.git] / 2023-hardware / notes.txt

diff --git a/2023-hardware/notes.txt b/2023-hardware/notes.txt
new file mode 100644 (file)
index 0000000..a510d65
--- /dev/null
+++ b/2023-hardware/notes.txt
@@ -0,0 +1,184 @@
+## Agenda
+
+I've prepared some information about
+
+ * current RISC-V hardware
+ * future RISC-V hardware
+ * performance numbers
+ * notes on QEMU
+
+
+## Koji
+
+http://fedora.riscv.rocks/koji/
+https://koji.fedoraproject.org/koji/
+
+Primary architecture or koji-shadow?
+
+
+## Current RISC-V hardware
+
+* Milk-V Pioneer (SG2042)
+* Sophgo 2U server (SG2042)
+
+  - Possible performance and quality issues
+  - Expensive
+  - Likely to end up as e-waste soon
+
+* SiFive Unmatched
+
+  - Slow
+  - Doesn't support any recent extensions, esp. V, H
+  - No BMC, but might use BMC-on-PCIe card
+
+* Single board computers (SBCs)
+
+  - VisionFive2 (StarFive JH7xxx), Lichee Pi 4A, ...
+  - Slow
+  - Limited memory, cores
+  - Reliability issues
+  - Upfront & ongoing engineering headaches integrating into a 19" rack
+
+* QEMU
+
+  - Performance numbers below, but not great
+  - Really using x86-64 hardware
+  - Hardware (x86-64) is well known and easy to manage
+  - No "e-waste", servers can be repurposed
+  - Supports all the latest extensions
+  - Supports any amount of RAM, large numbers of vCPUs
+  - We can add new extensions and fix bugs relatively easily
+
+
+
+## Future RISC-V hardware
+
+* Sophgo SG2380
+
+  - 16 x SiFive P670
+  - announced yesterday
+
+* SiFive P870
+
+  - announced in August 2023
+
+* StarFive JH8100
+
+  - TSMC 12nm
+  - H extension
+  - power and efficiency versions, but they are not fully compatible
+  - 8 lanes of gen3 PCIe
+  - 4 USB 3.2 gen2
+  - miniITX development board
+
+* Ventana
+
+* Rivos
+
+
+## Performance numbers
+
+
+               binutils     openssl      python3.12   mingw-gcc
+
+i686           1589         1577         3411         4292
+
+x86-64         1419         1172         2462         1827
+
+aarch64        1573          811 (?)     1845         2521
+
+ppc64le        2165         1291         3073         4388
+
+s390x          2553         1380         1984 (?)     6824
+
+
+
+qemu-system-riscv64 16 vCPUs, 16 GB
+on AMD Ryzen 9 7950X server
+
+  (LTO)        4493         3052        14502        12428
+                 +217%        +160%       +489%        +580%
+
+  (no LTO)     3267         1351         6353        (failed)
+                 +130%        +15%        +158%
+
+qemu-system-riscv64 32 vCPUs, 16 GB
+on AMD Genoa-X server
+
+  (LTO)        6841         3182
+
+  (no LTO)     5115         1882         (failed)
+
+
+VisionFive 2
+  (LTO)        7202         8823                     (crashed in LTO step)
+                 +408%        +653%
+
+  (no LTO)     3274         2059         11627
+                 +130%        +75%
+
+
+## Single thread performance
+
+qemu-system-riscv64   912
+on AMD Genoa-X
+
+HiFive Unmatched      616
+
+qemu-system-riscv64   598
+on AMD 7950x
+
+VisionFive 2          425
+
+Koji/ppc64le          144
+
+Koji/i686             105
+
+Koji/x86-64           100
+
+Koji/aarch64           89
+
+Koji/s390x             65
+
+AMD Genoa-X (x86-64)   36
+
+AMD 7950x (x86-64)     35
+
+
+
+## QEMU observations
+
+"TCG" is the name for QEMU's software emulation, eg. RISC-V fully
+emulated guest on x86-64 host.
+
+Works using Translation Blocks (TBs) which translate basic blocks of
+guest code.
+
+Well understood (by me), easy to fix simpler issues.
+
+ - I posted a patch yesterday which gets ~ +6% performance gain
+
+A few tips to make TCG run (a bit) fast(er):
+
+ - Compile with -march=native (+4%)
+
+ - Profile with perf
+
+ - Don't overprovision host CPUs
+
+    * However pinning vCPUs to pCPUs didn't really help
+
+ - Give it plenty of guest & host RAM
+
+    * Measured memory overhead on host is up to 40% after running
+      for some time
+
+    * Host TBs track guest page cache; as long as a translated
+      executable remains in the guest page cache, it will not be
+      retranslated
+
+ - Don't restart the VM
+
+ - Software TLB
+
+ - Fast vs slow jumps