docs(journal): add slide about t-test and more data for 512 case

doc/demo-journal/slides-journal.md

@@ -224,18 +224,48 @@ done
 echo 1 | sudo tee /sys/devices/system/cpu/intel_pstate/no_turbo
 
 # The following run is just a (CPU) warm up, the results are discarded.
-cabal run bench-sqlite
+bench-sqlite
 
 for i in $(seq 10); do
   for j in $(seq 2 12); do
-    cabal run bench-journal -- $((2**$j)) >> /tmp/bench-journal-$j.txt
-    cabal run bench-sqlite  -- $((2**$j)) >> /tmp/bench-sqlite-$j.txt
+    bench-journal -- $((2**$j)) >> /tmp/bench-journal-$j.txt
+    bench-sqlite  -- $((2**$j)) >> /tmp/bench-sqlite-$j.txt
   done
 done
 ```
 
 # Benchmark statistics
 
+```R
+Input512=("
+  workload                   throughput
+  journal-512                6129.90
+  ...
+  journal-512                5756.72
+  sqlite-512                 4851.46
+  ...
+  sqlite-512                 6735.40
+")
+
+df512 = read.table(textConnection(Input512),header=TRUE)
+bartlett.test(df512$throughput ~ df512$workload, data=df512)
+
+t.test(df512$throughput ~ df512$workload, data=df512,
+       var.equal=TRUE,
+       conf.level=0.95)
+```
+
+* Output
+
+```bash
+data:  df512$throughput by df512$workload
+t = 0.29646, df = 18, p-value = 0.7703
+mean in group journal-512  mean in group sqlite-512
+                 5937.373                  5839.218
+```
+
+# Benchmark plot
+
 ![](./images/journal-benchmark.jpg)
 
 # Amdahl's law vs the Universal scalability law
@@ -265,7 +295,7 @@ done
 # Summary
 
 * We have shown how to use the journal to:
-  - Faster write path than with a database (lock-free and append only)
+  - Faster write path than with a database (lock-free, append only and zero-copy)
   - Get faster crash recovery for free
   - Get all the deterministic testing stuff for free
   - Rich time traveling debugger
@@ -298,6 +328,9 @@ done
   for the journal version of the service;
 
 * Using the Linux kernel's `io_uring` to ammortise the cost of syscalls (by
-  batching and doing them async).
+  batching and doing them async);
+
+* But protocol improvements are likely much more important for performance than
+  these low-level database engine and event loop improvements!
 
 # Thanks! Questions? References:

src/journal/src/Journal/Internal/BufferClaim.hs

@@ -42,11 +42,15 @@ commit :: BufferClaim -> Logger -> IO ()
 commit (BufferClaim bb) logger = do
   let Capacity frameLen = getCapacity bb
   logg logger ("commit, frameLen: " ++ show frameLen)
+  -- XXX:
+  -- forceAt bb 0 frameLen
   writeFrameType bb 0 Valid
   writeFrameLength bb 0 (HeaderLength (int2Int32 frameLen))
+{-# INLINE commit #-}
 
 abort :: BufferClaim -> IO ()
 abort (BufferClaim bb) = do
   let Capacity frameLen = getCapacity bb
   writeFrameType bb 0 Padding
   writeFrameLength bb 0 (HeaderLength (int2Int32 frameLen))
+{-# INLINE abort #-}

src/journal/src/Journal/Internal/ByteBufferPtr.hs

@@ -635,11 +635,14 @@ force :: ByteBuffer -> IO ()
 force bb =
   withForeignPtr (bbData bb) $ \ptr ->
     msync ptr (fromIntegral (bbCapacity bb)) mS_SYNC False
+{-# INLINE force #-}
 
 forceAt :: ByteBuffer -> Int -> Int -> IO ()
-forceAt bb offset len =
+forceAt bb offset len = do
+  Slice slice <- readIORef (bbSlice bb)
   withForeignPtr (bbData bb) $ \ptr ->
-    msync (ptr `plusPtr` offset) (fromIntegral len) mS_SYNC False
+    msync (ptr `plusPtr` (slice + offset)) (fromIntegral len) mS_SYNC False
+{-# INLINE forceAt #-}
 
 ------------------------------------------------------------------------
 

src/sut/dumblog/bench/Common.hs

@@ -42,7 +42,7 @@ rEAD_FREQUENCY :: Int
 rEAD_FREQUENCY = 80
 
 iTERATIONS :: Int
-iTERATIONS = 50000
+iTERATIONS = 100000
 
 vALUE_TO_WRITE :: ByteString
 vALUE_TO_WRITE = LBS.pack "Dumblog"

src/sut/dumblog/tools/benchmark-plot.r

@@ -240,7 +240,7 @@ df256 = read.table(textConnection(Input256),header=TRUE)
 bartlett.test(df256$throughput ~ df256$workload, data=df256)
 
 t.test(df256$throughput ~ df256$workload, data=df256,
-       var.equal=TRUE,
+       var.equal=FALSE,
        conf.level=0.95)
 
 Input512=("
@@ -255,6 +255,26 @@ journal-512                5780.60
 journal-512                5756.72
 journal-512                6042.59
 journal-512                5895.52
+journal-512                5901.41
+journal-512                5850.24
+journal-512                5947.85
+journal-512                5907.77
+journal-512                6093.77
+journal-512                5928.50
+journal-512                5827.59
+journal-512                5982.38
+journal-512                6003.93
+journal-512                5882.08
+journal-512                5875.74
+journal-512                5744.34
+journal-512                5643.23
+journal-512                6082.79
+journal-512                6206.22
+journal-512                5982.41
+journal-512                5722.13
+journal-512                5902.64
+journal-512                5924.12
+journal-512                5849.07
 sqlite-512               4851.46
 sqlite-512               4175.24
 sqlite-512               5001.94
@@ -265,6 +285,26 @@ sqlite-512               6603.62
 sqlite-512               6625.08
 sqlite-512               6482.68
 sqlite-512               6430.65
+sqlite-512               4714.32
+sqlite-512               4250.24
+sqlite-512               6604.46
+sqlite-512               4997.60
+sqlite-512               4718.35
+sqlite-512               6752.57
+sqlite-512               4274.27
+sqlite-512               6255.25
+sqlite-512               5831.62
+sqlite-512               4677.35
+sqlite-512               6265.19
+sqlite-512               4534.84
+sqlite-512               4561.84
+sqlite-512               4579.69
+sqlite-512               6674.64
+sqlite-512               4278.95
+sqlite-512               6830.67
+sqlite-512               6454.88
+sqlite-512               5122.86
+sqlite-512               4319.10
 ")
 
 df512 = read.table(textConnection(Input512),header=TRUE)
@@ -302,7 +342,7 @@ df1024 = read.table(textConnection(Input1024),header=TRUE)
 bartlett.test(df1024$throughput ~ df1024$workload, data=df1024)
 
 t.test(df1024$throughput ~ df1024$workload, data=df1024,
-       var.equal=TRUE,
+       var.equal=FALSE,
        conf.level=0.95)
 
 Input2048=("
@@ -374,7 +414,7 @@ c32   <- c(6440.925, 6015.705)
 c64   <- c(6238.401, 6054.991)
 c128  <- c(6129.595, 6831.315)
 c256  <- c(6179.770, 6013.608)
-c512  <- c(5937.373, 5839.218)
+c512  <- c(5921.065, 5503.029)
 c1024 <- c(5828.134, 4248.163)
 c2048 <- c(5444.993, 3802.999)
 c4096 <- c(4924.328, 3549.512)