Thursday, July 2, 2015

An Exercise in Profiling a Go Program

Recently, while working on my current project ogonori, a Go client for the OrientDB database, I found that I had a defect in the code that encodes and decodes integer values in the way that the OrientDB binary network protocol requires (namely zigzag encoding, followed by encoding that output as a variable length integer).

After fixing the issue, first with the encoder and then with the decoder, I decided that I should do an exhaustive test of all 64 bit integers: start with MinInt64 (-9223372036854775808), zigzag-encode it, varint encode it, then varint decode it and zigzag-decode it and you should get back the number you started with. Increment by 1 and try it again, until you reach MaxInt64 (9223372036854775807).

(Note: I only have to use the Min/Max range of signed integers, since OrientDB is a Java database and only allows signed ints.)

I ran a small range of the possible 64-bit integer space and found that doing this exhaustive test was going to take a very long time. Since I have 8 CPUs on my system, I decided to first parallelize the test into 8 separate goroutines, each taking 1/8 of the total range:

With this code, I spawn 8 threads, running 10 goroutines. Eight of them do the encoding/decoding test and if any integer encode/decode fails the test it is written to "failure channel" of type chan string, which the main goroutine monitors.

A sync.WaitGroup (a counting semaphore) is created and shared among the goroutines. When each "range tester" finishes, it calls Done() on the WaitGroup to decrement the semaphore. The final (nameless) goroutine waits until all "range tester" goroutines have finished and then closes the single shared failure channel.

Closing of the failure channel, causes the loop over that channel in the main goroutine to exit and the whole program finishes.

/* ---[ Performance Baseline ]--- */

I fired this up with the following smaller testrange:

    ranges := []testrange{
        {100000001, 150000001},
        {200000001, 250000001},
        {300000001, 350000000},
        {400000001, 450000000},
        {500000001, 550000000},
        {600000001, 650000000},
        {700000001, 750000000},
        {800000001, 850000000},

and ran top. To my surprise I was only using ~400% CPU, rather than ~800% (the max my system supports):

$ top -d1
 PID USER      PR  ... S  %CPU %MEM     TIME+ COMMAND                                                       
1736 midpete+  20  ... S 420.9  0.0   1:31.33 ogonori

I then looked at the CPU usage of each thread using the -H option to top and saw that my 8 range-tester goroutines were each using only about 50% CPU. And that there was a 9th thread that was also consistently using 40 to 50% CPU. My guess was that this was a GC thread.

$ top -d1 -H
 PID USER      PR  ...  S %CPU %MEM     TIME+ COMMAND                                                        
1740 midpete+  20  ...  S 50.1  0.0   0:21.47 ogonori                                                        
1744 midpete+  20  ...  R 50.1  0.0   0:21.52 ogonori                                                        
1742 midpete+  20  ...  S 49.2  0.0   0:21.38 ogonori                                                        
1736 midpete+  20  ...  S 47.2  0.0   0:21.53 ogonori                                                        
1738 midpete+  20  ...  S 46.2  0.0   0:22.11 ogonori                                                        
1745 midpete+  20  ...  R 46.2  0.0   0:20.37 ogonori                                                        
1741 midpete+  20  ...  S 45.2  0.0   0:21.41 ogonori                                                        
1743 midpete+  20  ...  R 42.3  0.0   0:21.26 ogonori                                                        
1739 midpete+  20  ...  S 40.3  0.0   0:21.35 ogonori                                                        
1737 midpete+  20  ...  S  3.9  0.0   0:02.07 ogonori                  

So I have an algorithm that should be trivially parallelizable with no shared memory and no contention (in theory), but it was only using half the CPU available to it. Hmmm...

Next I ran the test on my system several times to get a baseline performance metric:

$ time ./ogonori -z  # the -z switch tells the ogonori code to only this
                     # benchmark rather than the usual OrientDB tests

Run1: real    3m44.602s
Run2: real    3m42.818s
Run3: real    3m28.917s

Avg ± StdErr:  218.8 ± 5 sec

Then I remembered I had not turned off the CPU power saving throttling on my Linux system (it was set to ondemand), so I ran the following script and repeated the benchmarks:

for i in /sys/devices/system/cpu/cpu[0-7]
    echo performance > $i/cpufreq/scaling_governor

$ time ./ogonori -z
Run1: real    2m12.605s
Run2: real    2m12.382s
Run3: real    2m13.172s
Run4: real    2m18.992s
Run5: real    2m17.538s
Run6: real    2m14.437s

Avg ± StdErr:  134.9 ± 1 sec

Wow, OK. So that alone gave me about a 60% improvement in throughput. Off to a good start.

/* ---[ Profiling the Code ]--- */

If you've never read Russ Cox's 2011 blog post on profiling a Go program, put it on your list - it is a treat to read.

Using what I learned there, I profiled the zigzagExhaustiveTest code to see how and where to improve it.

$ ./ogonori -z

I then opened the .prof file with golang's pprof tool and looked at the top 10 most heavily used functions:

$ rlwrap go tool pprof ogonori
  # Using rlwrap gives you bash-like behavior and history

(pprof) top 10
171.48s of 255.92s total (67.01%)
Dropped 171 nodes (cum <= 1.28s)
Showing top 10 nodes out of 36 (cum >= 8.78s)
      flat  flat%   sum%        cum   cum%
    45.98s 17.97% 17.97%     45.98s 17.97%  scanblock
    25.63s 10.01% 27.98%     33.58s 13.12%  runtime.mallocgc
    19.20s  7.50% 35.48%    111.35s 43.51%  g/q/o/o/b/varint.ReadVarIntToUint
    14.94s  5.84% 41.32%     15.62s  6.10%  bytes.(*Buffer).grow
    12.44s  4.86% 46.18%     12.44s  4.86%  runtime.MSpan_Sweep
    11.87s  4.64% 50.82%     15.93s  6.22%  bytes.(*Buffer).Read
    11.33s  4.43% 55.25%     21.56s  8.42%  bytes.(*Buffer).WriteByte
    11.18s  4.37% 59.62%     11.18s  4.37%  runtime.futex
    10.13s  3.96% 63.57%     19.16s  7.49%  bytes.(*Buffer).Write
     8.78s  3.43% 67.01%      8.78s  3.43%  runtime.memmove

(pprof) top10 -cum
110.32s of 255.92s total (43.11%)
Dropped 171 nodes (cum <= 1.28s)
Showing top 10 nodes out of 36 (cum >= 25.50s)
      flat  flat%   sum%        cum   cum%
         0     0%     0%    147.62s 57.68%  runtime.goexit
     2.94s  1.15%  1.15%    147.49s 57.63%  main.func·018
    19.20s  7.50%  8.65%    111.35s 43.51%  g/q/o/o/b/varint.ReadVarIntToUint
         0     0%  8.65%     77.81s 30.40%  GC
    45.98s 17.97% 26.62%     45.98s 17.97%  scanblock
     4.90s  1.91% 28.53%     38.48s 15.04%  runtime.newobject
    25.63s 10.01% 38.55%     33.58s 13.12%  runtime.mallocgc
     6.65s  2.60% 41.15%     31.39s 12.27%  g/q/o/o/b/varint.VarintEncode
         0     0% 41.15%     30.48s 11.91%  System
     5.02s  1.96% 43.11%     25.50s  9.96%  encoding/binary.Read

We can see that a significant percentage of time (>30%) is being spent in GC, so the program is generating a lot of garbage somewhere - plus the cost of generating new heap data, which the runtime.mallocgc figure tells me is at least 13% of the program run time.

Remember that there are four steps to my algorithm:

  1. zigzag encode (varint.ZigzagEncodeUInt64)
  2. varint encode (varint.VarintEncode)
  3. varint decode (varint.ReadVarIntToUint)
  4. zigzag decode (varint.ZigzagDecodeInt64)

The zigzag encode/decode steps are simple bit manipulations, so they are fast. Typing web at the pprof prompt launches an SVG graph of where time was spent. The zigzag functions don't even show up - they were dropped off as being too small (not shown here).

So I needed to focus on steps 2 and 3 which take (cumulatively) 43.5% and 12.3%, respectively.

Since varint.ReadVarIntToUint is the biggest offender let's look at it in detail in the pprof tool:

I've marked the biggest time sinks with an arrow on the left side. Generally one should start with the biggest bottleneck, so let's rank these by cumulative time (2nd col):

->  32.41s   111:   err = binary.Read(&buf, binary.LittleEndian, &u)
->  16.83s    73:   n, err = r.Read(ba[:])
->  15.93s   106:   buf.WriteByte(y | z)
->  14.82s    88:   var buf bytes.Buffer
->   8.53s   110:   padTo8Bytes(&buf)

First, it is very interesting how expensive creating a bytes.Buffer is. But first we need to deal with binary.Read.

Because I'm only ever passing in uint64's, the only real functionality I'm using in this function is:

*data = order.Uint64(bs)

/* ---[ Optimization #1 ]--- */

But it's even worse. If you look back at varint.ReadVarIntToUint you'll see that I'm creating a bytes.Buffer and copying bytes into it only so that I can pass that Buffer (as an io.Reader) into the binary.Read function:

err = binary.Read(buf, binary.LittleEndian, &u)

which then immediately copies all those bytes back out of the buffer:

if _, err := io.ReadFull(r, bs); err != nil {
    return err

So this is nothing but wasteful data copying and the heap allocations for it.

binary.Read also does a type switch where a good percentage of time is spent

     2.01s      4.49s    151:       switch data := data.(type) {

and, as stated, the only useful method ever called in it is:

 -->  460ms      2.76s    167:          *data = order.Uint64(bs)

So I should try just calling binary.LittleEndian.Uint64(bs) directly.

Here's the revised varint.ReadVarIntToUint function (with everything inlined for easier reading and profiling analysis):

This change also removes the padTo8Bytes method that wrote one byte at a time to the bytes.Buffer and took >3% of program time itself.

Now let's rerun the benchmarks:

Run 1: real  0m27.182s
Run 2: real  0m27.053s
Run 3: real  0m28.200s
Run 4: real  0m25.762s
Run 5: real  0m26.031s
Run 6: real  0m26.813s

Avg ± StdErr:  26.8 ± 0.4 sec

Outstanding! Throughput increased 5x (134.9/26.8). And using top, I see that the goroutines are consuming nearly all available CPU:

$ top -d1
12983 midpete+  20   0  352496   5768   2736 R 763.7  0.0   1:35.64 ogonori

$ top -d1 -H
  PID USER      PR  NI    VIRT    RES    SHR S %CPU %MEM     TIME+ COMMAND                                                       
13231 midpete+  20   0  286960   5772   2744 R 97.5  0.0   0:22.51 ogonori                                                       
13225 midpete+  20   0  286960   5772   2744 R 91.7  0.0   0:22.47 ogonori                                                       
13227 midpete+  20   0  286960   5772   2744 R 90.7  0.0   0:23.09 ogonori                                                       
13232 midpete+  20   0  286960   5772   2744 S 90.7  0.0   0:22.26 ogonori                                                       
13235 midpete+  20   0  286960   5772   2744 R 90.7  0.0   0:09.72 ogonori                                                       
13230 midpete+  20   0  286960   5772   2744 R 88.7  0.0   0:22.14 ogonori                                                       
13233 midpete+  20   0  286960   5772   2744 R 73.1  0.0   0:22.70 ogonori                                                       
13228 midpete+  20   0  286960   5772   2744 R 71.2  0.0   0:22.39 ogonori                                                       
13229 midpete+  20   0  286960   5772   2744 R 70.2  0.0   0:23.09 ogonori 

I also used pprof to profile this run, so let's examine compare the cumulative top10 before and after:

Before (reprinted from above):

(pprof) top10 -cum
110.32s of 255.92s total (43.11%)
Dropped 171 nodes (cum <= 1.28s)
Showing top 10 nodes out of 36 (cum >= 25.50s)
      flat  flat%   sum%        cum   cum%
         0     0%     0%    147.62s 57.68%  runtime.goexit
     2.94s  1.15%  1.15%    147.49s 57.63%  main.func·018
    19.20s  7.50%  8.65%    111.35s 43.51%  g/q/o/o/b/varint.ReadVarIntToUint
         0     0%  8.65%     77.81s 30.40%  GC
    45.98s 17.97% 26.62%     45.98s 17.97%  scanblock
     4.90s  1.91% 28.53%     38.48s 15.04%  runtime.newobject
    25.63s 10.01% 38.55%     33.58s 13.12%  runtime.mallocgc
     6.65s  2.60% 41.15%     31.39s 12.27%  g/q/o/o/b/varint.VarintEncode
         0     0% 41.15%     30.48s 11.91%  System
     5.02s  1.96% 43.11%     25.50s  9.96%  encoding/binary.Read


(pprof) top15 -cum
63680ms of 65970ms total (96.53%)
Dropped 33 nodes (cum <= 329.85ms)
Showing top 15 nodes out of 18 (cum >= 930ms)
      flat  flat%   sum%        cum   cum%
    2280ms  3.46%  3.46%    64470ms 97.73%  main.func·018
         0     0%  3.46%    64470ms 97.73%  runtime.goexit
   17760ms 26.92% 30.38%    34190ms 51.83%  g/q/o/o/b/varint.ReadVarIntToUint
    5890ms  8.93% 39.31%    26370ms 39.97%  g/q/o/o/b/varint.VarintEncode
    8550ms 12.96% 52.27%    16360ms 24.80%  bytes.(*Buffer).Write
    9080ms 13.76% 66.03%    11500ms 17.43%  bytes.(*Buffer).Read
    1460ms  2.21% 68.24%     7550ms 11.44%  runtime.newobject
    4370ms  6.62% 74.87%     6090ms  9.23%  runtime.mallocgc
    5650ms  8.56% 83.43%     5650ms  8.56%  runtime.memmove
    4580ms  6.94% 90.37%     4580ms  6.94%  bytes.(*Buffer).grow
     680ms  1.03% 91.41%     1630ms  2.47%  bytes.(*Buffer).Reset
    1500ms  2.27% 93.68%     1500ms  2.27%  encoding/binary.littleEndian.Uint64
         0     0% 93.68%     1030ms  1.56%  GC
     950ms  1.44% 95.12%      950ms  1.44%  bytes.(*Buffer).Truncate
     930ms  1.41% 96.53%      930ms  1.41%  runtime.gomcache

More good news. In the previous version, GC was taking 30% of the total CPU time. Now, more than 90% of the time is now being spent in the two main workhorse methods: varint.ReadVarIntToUint and varint.VarintEncode. GC time has been reduced to 1.5%!

I suspect the reason that goroutines in the earlier code version only took 40-50% of a CPU is because GC was the contention point. Garbage Collection in golang is a stop-the-world affair, so all other threads are paused until it finishes. By reducing GC to only 1.5%, now the range-testing goroutines can spend far more time running - approaching 100%.

/* ---[ Optimization #2 ]--- */

Are there further improvements we can make? Since the program now spends 40% of its time in varint.VarintEncode, let's look at that function in detail:

Almost 75% of the time in this function is spent writing to the io.Writer (a bytes.Buffer). We write one byte at a time to it. Perhaps it would be better to write it all to a byte slice first and then issue one w.Write.

The new code is then:

And the next round of benchmarks are:

real    0m38.899s
real    0m45.135s
real    0m38.047s
real    0m42.377s
real    0m32.894s
real    0m37.962s
real    0m38.926s
real    0m37.870s

Avg ± StdErr: 39.0 ± 1.2

Hmm, not good. It looks like this second revision caused my code to go backwards in performance by 30%. To be sure, I reverted the change and re-ran the benchmarks with only optimization #1 again: they returned to the ~25s/run timeframe I saw before. So it is true: this second change made things worse.

And the analysis of top agreed: the goroutines were no long using 90%+ CPU:

$ top -d1
  PID USER      PR  NI    VIRT    RES    SHR S  %CPU %MEM     TIME+ COMMAND                                                      
22149 midpete+  20   0  286960   5776   2744 R 593.9  0.0   1:06.66 ogonori 

$ top -d1 -H
  PID USER      PR  NI    VIRT    RES    SHR S %CPU %MEM     TIME+ COMMAND                                                       
22205 midpete+  20   0  229620   7812   2744 R 74.6  0.0   0:10.68 ogonori                                                       
22201 midpete+  20   0  229620   7812   2744 R 73.6  0.0   0:10.14 ogonori                                                       
22202 midpete+  20   0  229620   7812   2744 S 71.6  0.0   0:10.77 ogonori                                                       
22207 midpete+  20   0  229620   7812   2744 S 70.6  0.0   0:10.97 ogonori                                                       
22206 midpete+  20   0  229620   7812   2744 R 68.7  0.0   0:11.14 ogonori                                                       
22204 midpete+  20   0  229620   7812   2744 R 65.7  0.0   0:10.07 ogonori                                                       
22199 midpete+  20   0  229620   7812   2744 R 56.9  0.0   0:10.98 ogonori                                                       
22203 midpete+  20   0  229620   7812   2744 R 53.0  0.0   0:11.19 ogonori                                                       
22197 midpete+  20   0  229620   7812   2744 R 43.2  0.0   0:11.17 ogonori                                                       
22200 midpete+  20   0  229620   7812   2744 S 17.7  0.0   0:09.95 ogonori                                                       
22198 midpete+  20   0  229620   7812   2744 S  3.9  0.0   0:00.77 ogonori         

Let's look at the pprof data for revision #2:

(pprof) top10 -cum
59.31s of 86.44s total (68.61%)
Dropped 92 nodes (cum <= 0.43s)
Showing top 10 nodes out of 24 (cum >= 5.36s)
      flat  flat%   sum%        cum   cum%
     1.95s  2.26%  2.26%     74.02s 85.63%  main.func·018
         0     0%  2.26%     74.02s 85.63%  runtime.goexit
    20.71s 23.96% 26.21%     44.80s 51.83%  g/q/o/o/b/varint.ReadVarIntToUint
     5.57s  6.44% 32.66%     24.39s 28.22%  g/q/o/o/b/varint.VarintEncode
       15s 17.35% 50.01%     18.71s 21.65%  bytes.(*Buffer).Read
     5.64s  6.52% 56.54%     13.46s 15.57%  runtime.makeslice
         0     0% 56.54%      8.39s  9.71%  GC
     5.86s  6.78% 63.32%      8.23s  9.52%  runtime.mallocgc
     2.48s  2.87% 66.18%      7.82s  9.05%  runtime.newarray
     2.10s  2.43% 68.61%      5.36s  6.20%  bytes.(*Buffer).Write

Now GC is back up to nearly 10% of the total running time. So let's look at the profile of the VarintEncode function we changed:

(pprof) list VarintEncode
Total: 1.44mins
   5.57s     24.39s (flat, cum) 28.22% of Total
       .          .     40://
   290ms      290ms     41:func VarintEncode(w io.Writer, v uint64) error {
   550ms     14.01s     42: bs := make([]byte, 0, 10)
   170ms      170ms     43: for (v & 0xffffffffffffff80) != 0 {
   2.04s      2.04s     44:     bs = append(bs, byte((v&0x7f)|0x80))
   320ms      320ms     45:     v >>= 7
       .          .     46: }
   680ms      680ms     47: bs = append(bs, byte(v&0x7f))
       .          .     48:
   1.20s      6.56s     49: n, err := w.Write(bs)
   120ms      120ms     50: if err != nil {
       .          .     51:     return oerror.NewTrace(err)
       .          .     52: }
       .          .     53: if n != len(bs) {
       .          .     54:     return fmt.Errorf("Incorrect number of bytes written. Expected %d. Actual %d", len(bs), n)
       .          .     55: }
   200ms      200ms     56: return nil
       .          .     57:}

We can see that 58% of the time of this method is spent allocating new memory (the []byte slice on line 42), thereby causing GC to take longer. Here's why - if you look at the implementation of bytes.Buffer, you'll see that it has a fixed bootstrap array it allocates to handle small buffers and another fixed byte array (runeBytes) to handle writes to WriteByte; both of these allow it to avoid memory allocation for small operations.

Since my test code is reusing the same bytes.Buffer for each iteration, no new allocations were occurring during each call to varint.VarintEncode. But with this second revision I'm creating a new byte slice of capacity 10 in each round. So this change should be reverted.

/* ---[ Lessons Learned ]--- */

When you have an algorithm that you think should be CPU bound and your threads are not using ~100% CPU, then you have contention somewhere. In many scenarios that will be IO wait. But if you have no IO in that portion of your app, then you either have hidden thread contention (mutexes) and/or you may have a lot of garbage collection happening, which pauses all your worker threads/goroutines while GC is happening. Use the pprof tool to determine where time is being spent.

For performance sensitive algorithms, you will want to be garbage free in the main path as much as possible.

Once you know where the time is going, you should generally go after the largest bottleneck first. There's always a primary bottleneck somewhere. Removing the bottleneck in one place causes it to move to another. In my case, I wanted that bottleneck to just be CPU speed (or as is often the case, the time to get data from main memory or a CPU cache into a register).

A big lesson learned here is to be wary of convenience methods in Go's standard library. Many are provided for convenience, not performance. The binary.Read(buf, binary.LittleEndian, &u) call in my case is one such example. The third parameter to binary.Read is of type interface{}, so a type switch has to be done to detect the type. If your code is only ever passing in one type (uint64 in my case), then go read the stdlib code and figure out if there is a more direct method to call. That change contributed to a 5x throughput improvement in my case!

Next, be careful of too much data copying. While the io.Writer is a nice interface, if you are working with byte slices and want to pass it to some stdlib method that requires io.Writer, you will often copy the data into a bytes.Buffer and pass that in. If the function you call copies those bytes back out to yet another byte slice, then garbage is being generated and time is being wasted. So be aware of what's happening in the methods you call.

Finally, always measure carefully before and after any attempted optimizations. Intuition about where bottlenecks are and what will speed things up are often wrong. The only thing of value is to measure objectively. To end I'll quote "Commander" Pike:

Rule 1. You can't tell where a program is going to spend its time. Bottlenecks occur in surprising places, so don't try to second guess and put in a speed hack until you've proven that's where the bottleneck is. --Rob Pike's 5 Rules of Programming

/* ---[ Misc Appendix Notes ]--- */

The overall int64 space will still take too long to run even with these improvements, so I've settled for sampling from the state space instead.

All benchmark comparisons done were statistically significant (p<0.01) using Student's t-test, as analyzed with this tool: The mean and standard errors were also calculated here.

I notice that even with my best optimization (#1), there is still a ninth thread using >70% CPU. I used kill -QUIT on the program to get a stack dump of all the goroutines. I get 10 go routines - the 8 doing the fnRangeTester work, one waiting on the WaitGroup and the main goroutine which is waiting on the range failchan line. So I'm not sure what that 9th thread is doing churning up 50-70% CPU. Anyone know how to tell?

[Update - 08-July-2015]

In the comments, Carlos Torres asked for the pprof line-by-line output of the ReadVarIntToUint function after the first optimization. I did two profiling runs and compared the pprof outputs and they were both nearly identical. Here is one of them:

(pprof) list ReadVarIntToUint
Total: 1.13mins
ROUTINE ======================== g/q/o/o/b/varint.ReadVarIntToUint 
 18.20s     35.45s (flat, cum) 52.08% of Total
      .          .     25://
  480ms      480ms     26:func ReadVarIntToUint(r io.Reader) (uint64, error) {
      .          .     27:    var (
  270ms      270ms     28:        varbs []byte
  120ms      3.84s     29:        ba    [1]byte
      .          .     30:        u     uint64
      .          .     31:        n     int
  180ms      180ms     32:        err   error
      .          .     33:    )
      .          .     34:
  260ms      260ms     35:    varbs = make([]byte, 0, 10)
      .          .     36:
      .          .     37:    /* ---[ read in all varint bytes ]--- */
      .          .     38:    for {
  3.84s     15.72s     39:        n, err = r.Read(ba[:])
  530ms      530ms     40:        if err != nil {
      .          .     41:            return 0, oerror.NewTrace(err)
      .          .     42:        }
   10ms       10ms     43:        if n != 1 {
      .          .     44:            return 0, oerror.IncorrectNetworkRead{Expected: 1, Actual: n}
      .          .     45:        }
  3.21s      3.21s     46:        varbs = append(varbs, ba[0])
  980ms      980ms     47:        if IsFinalVarIntByte(ba[0]) {
  570ms      570ms     48:            varbs = append(varbs, byte(0x0))
      .          .     49:            break
      .          .     50:        }
      .          .     51:    }
      .          .     52:
      .          .     53:    /* ---[ decode ]--- */
      .          .     54:
      .          .     55:    var right, left uint
      .          .     56:
  620ms      620ms     57:    finalbs := make([]byte, 8)
      .          .     58:
      .          .     59:    idx := 0
  1.08s      1.08s     60:    for i := 0; i < len(varbs)-1; i++ {
  360ms      360ms     61:        right = uint(i) % 8
   20ms       20ms     62:        left = 7 - right
  230ms      230ms     63:        if i == 7 {
      .          .     64:            continue
      .          .     65:        }
  840ms      840ms     66:        vbcurr := varbs[i]
  900ms      900ms     67:        vbnext := varbs[i+1]
      .          .     68:
  120ms      120ms     69:        x := vbcurr & byte(0x7f)
  670ms      670ms     70:        y := x >> right
  670ms      670ms     71:        z := vbnext << left
  650ms      650ms     72:        finalbs[idx] = y | z
  780ms      780ms     73:        idx++
      .          .     74:    }
      .          .     75:
  540ms      2.19s     76:    u = binary.LittleEndian.Uint64(finalbs)
  270ms      270ms     77:    return u, nil
      .          .     78:}

If you compare it to the pprof before the optimization, the top half looks about the same, but the bottom half is dramatically different. For example, more than 30s was spent in binary.Read(&buf, binary.LittleEndian, &u) in the original version. The replacement code, binary.LittleEndian.Uint64(finalbs), only takes up about 2 seconds of processing time.

The only remaining spot I see for any further optimization is the 15s (out of 35s) spent in r.Read(ba[:]). The problem, however, is that with a varint you don't know how many bytes long it is in advance, so you have read and examine them one at a time. There is probably a way to optimize this, but I haven't attempted it yet.


  1. hi there, nice writeup.

    If you can take stack dumps of the threads frequently, the 9th thread's responsibility will be known soon.

  2. If you set GODEBUG you can get regular state dumps from the scheduler to see what goroutines are running

  3. After the improvement was made on varint.ReadVarIntToUint, how did its detailed profile look in pprof?

    1. I've posted it in an update. Thanks for your interest!

    2. Thanks for the update.

      There are 2 things I find interesting here.

      1) I find surprising that it is so expensive (3.84s) to allocate a byte array of len 1;

      120ms 3.84s 29: ba [1]byte


      2) On line:

      3.84s 15.72s 39: n, err = r.Read(ba[:])

      I wonder how much time is attributed to just slicing the array, separate from the r.Read() call.

    3. > I find surprising that it is so expensive (3.84s) to allocate a byte array of len 1;

      Agreed. I'd like to understand this better. For example is the backing array is being created on the heap or the stack? If the heap, why? One of my performance frustrations with the io.Reader and io.Writer interfaces is that you always have to work with slices, never a single byte. And bytes.Buffer is a concrete type, not a interface.

  4. humbled and Andrew -

    Thanks for the suggestions!

    I've made a first attempt at understanding with the "ninth" thread is doing using stack dumps and setting the GODEBUG env var. The multiplexing of goroutines onto threads is not something I'm used from my JVM experience, so i'm still trying to get a handle on how to read the tea leaves. I'll post something if I make headway on it.

  5. Have you tried using the builtin functions

    1. Oh, man. It's amazing the gems that are present in the Go std library. Honestly, I never even thought to look for this there. But since the encoding is taken from protocol buffers, which is from Google and Go is from Google, it makes sense. Even so, I'm not sorry I did it on my own, just to have gotten back to doing some low-level bit fiddling, not something I get to do much at my day job :-).

      And I can now do a perf comparison between my version and theirs. I'll put it on my TODO list.

      Thanks for the heads-up!

  6. Very detailed article, I am going to start learning Go soon. bookmarking it. Thanks!! Paul

  7. Jadi, pastinya hal tersebut akan membuat para pemain bermain sampai lupa waktu, namun bermain sampai lupa waktu merupakan suatu hal yang tidak baik. Karena jika para pemain terus bermain seperti itu, maka para pemain akan mengalami kekalahan dalam taruhan slot ini.
    bandar ceme terbaik
    paito warna terlengkap
    forum prediksi

  8. Nice posting information I liked it
    Sanjary Kids is one of the best play school and preschool in Hyderabad,India. Give your child the best preschool experience by choosing the best playschool of Hyderabad in Abids. we provide programs like Play group,Nursery,Junior KG,Senior KG,and provides Teacher Training Program.
    Preschool teacher training course in hyderabad

  9. This is a very interesting reading value in the newspaper. In fact, I'm grateful to have given the chance to browse an informative article like this! In fact, I appreciate this article, thank you for sharing this information.

  10. Nice blog information provided by the author

    Sanjary Academy is the best Piping Design institute in Hyderabad, Telangana. It is the best Piping design Course in India and we have offer professional Engineering Courses like Piping design Course, QA/QC Course, document controller course, Pressure Vessel Design Course, Welding Inspector Course, Quality Management Course and Safety Officer Course.
    Piping Design Course
    Piping Design Course in Hyderabad ­
    Piping Design Course in India­

  11. nice post
    Yaaron Studios is one of the rapidly growing editing studios in Hyderabad. We are the best Video Editing services in Hyderabad. We provides best graphic works like logo reveals, corporate presentation Etc. And also we gives the best Outdoor/Indoor shoots and Ad Making services.
    Best video editing services in Hyderabad,ameerpet
    Best Graphic Designing services in Hyderabad,ameerpet­
    Best Ad Making services in Hyderabad,ameerpet­

  12. persentase kemenangan terbaik dan yang menawarkan harga terendah. Poker adalah salah satu permainan yang tergantung pada banyak faktor 98toto

  13. We develop free teaching aids for parents and educators to teach English to pre-school children. For more info please visit here: English for children

  14. Very detailed one to learn. Thanks will share it on our site.

  15. Incredible I love this!!!!!!!!!! hey all, I want to tell you that this is an awesome blog and which you have written in this blog, each and every word is also great. Nice dear!! Thanks to sharing with us! Thanks a lot, dear!!!!!!!!!
    India VPS Hosing

  16. Feeling good to read such a informative blog, mostly i eagerly search for this kind of blog. I really found your blog informative and unique, waiting for your new blog to read.
    Digital marketing Service in Delhi
    SMM Services
    PPC Services in Delhi
    Website Design & Development Packages
    SEO Services Packages
    Local SEO services
    E-mail marketing services
    YouTube plans

  17. This site helps to clear your all query.
    This is really worth reading. nice informative article.BA 3rd Year Time Table
    BA 2nd Year time table

  18. Congratulations, wonderful work, it's beautiful so much and good idea on site. CASINO ONLINE

  19. The counters, set up at 16 centres on the way to the four shrines, will record visitors' details —

    name, place, contact number (also of team leaders/tour operators), vehicle registration

    number, the destination shrine and the date of return. Their photographs will also be clicked.

    for Visit Here Char Dham Yatra package cost from Dehradun

  20. This comment has been removed by the author.

  21. The blog was useful in knowing about book life.Excellent blog thanks for sharing the valuable becomes easy to read and easily understand the information.Create you own project from us me project centers in chennai

  22. Akshi Engineers Pvt. Ltd. is leading Saudi Arabian Rolling Mill Manufacturers, innovating precision metal shaping solutions. Expertise in steel, aluminum, and specialty alloys. Committed to excellence in quality and efficiency for global industries.