When all samples are identical, instead of printing
+0.00%, print just 0.00%.

This is not just more accurate.
When benchmarking the compiler, one item of interest
is the impact on object file and executable sizes.
These can change by tiny amounts, below two sig figs.
However, the existence of any change is sometimes a cue
to the reader to investigate further.
This obviously should not be relied on,
but it is nevertheless a useful signal.

Change-Id: I1e5157a52fc47c6871ca87b0e9cff223f4fd43d9
Reviewed-on: https://go-review.googlesource.com/44190
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

compilebench emits both ns/op and user-ns/op measurements.
This CL improves benchstat's handling of such measurements.
Migrated from https://github.com/rsc/benchstat/pull/5.

Sample output, before:

name       old user-ns/op  new user-ns/op  delta
Template        279M ± 7%       273M ± 4%     ~     (p=0.329 n=6+5)
Unicode         148M ±11%       131M ± 7%  -11.49%  (p=0.030 n=6+5)
GoTypes         830M ± 2%       835M ± 3%     ~     (p=0.792 n=6+5)
SSA            9.60G ± 3%      9.64G ± 2%     ~     (p=0.792 n=6+5)
Flate           158M ± 3%       155M ± 1%     ~     (p=0.329 n=6+5)
GoParser        204M ± 6%       202M ± 8%     ~     (p=0.792 n=6+5)
Reflect         511M ± 9%       531M ± 6%     ~     (p=0.177 n=6+5)
Tar             151M ± 3%       151M ± 5%     ~     (p=0.931 n=6+5)
XML             271M ± 3%       268M ± 3%     ~     (p=0.429 n=6+5)

Sample output, after:

name       old user-time/op  new user-time/op  delta
Template         279ms ± 7%        273ms ± 4%     ~     (p=0.329 n=6+5)
Unicode          148ms ±11%        131ms ± 7%  -11.49%  (p=0.030 n=6+5)
GoTypes          830ms ± 2%        835ms ± 3%     ~     (p=0.792 n=6+5)
SSA              9.60s ± 3%        9.64s ± 2%     ~     (p=0.792 n=6+5)
Flate            158ms ± 3%        155ms ± 1%     ~     (p=0.329 n=6+5)
GoParser         204ms ± 6%        202ms ± 8%     ~     (p=0.792 n=6+5)
Reflect          511ms ± 9%        531ms ± 6%     ~     (p=0.177 n=6+5)
Tar              151ms ± 3%        151ms ± 5%     ~     (p=0.931 n=6+5)
XML              271ms ± 3%        268ms ± 3%     ~     (p=0.429 n=6+5)

Change-Id: I716b4e89a28adc72e9135b8c580434f748a9c0a9
Reviewed-on: https://go-review.googlesource.com/39794
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Quentin Smith <quentin@golang.org>

Before this change, localperf would log:

localperf: 2017/05/09 14:10:47 failed to fetch recent uploads: Get localhost:8081/uploads?extra_label=by&extra_label=upload-time&limit=16: unsupported protocol scheme "localhost"

Change-Id: Ifeba048c9dd54cf2c02ed539c5bfa0e6074fd60a
Reviewed-on: https://go-review.googlesource.com/43032Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

Change-Id: I18ee1d1eab704c628dd476ee3aa7dc90e729106d
Reviewed-on: https://go-review.googlesource.com/43053
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

Change-Id: I0ba5b534e0d5262e2626351d10564219d18f93a8
Reviewed-on: https://go-review.googlesource.com/41372Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

Change-Id: Ibbe3faf5bad9f702b46703a68c612b6e9dbf666d
Reviewed-on: https://go-review.googlesource.com/41250Reviewed-by: Russ Cox <rsc@golang.org>

This adds a -data flag to localperfdata to specify a local data
directory. In combination with -dsn to specify a local database,
localperfdata is now useful for non-testing applications.

Change-Id: I32ddb7ba5a96483c1f93cb7b52bbc2643b5b3798
Reviewed-on: https://go-review.googlesource.com/37861Reviewed-by: Russ Cox <rsc@golang.org>

MySQL requires HAVING to refer to a column from a subquery, but SQLite
requires WHERE since the query is not an aggregation. Plumb through
the driver name so ListUploads can change its query slightly.

Change-Id: I07bb8010824d7e2362e203a68fed7bb968d5fe40
Reviewed-on: https://go-review.googlesource.com/39614
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

Go 1.9 and up write "pkg", "goos", and "goarch" keys in benchmark
output. benchstat now understands benchmark labels, and uses them to
separate incomparable benchmark results. cmd/benchstat gains a
command-line flag called "-split" to control this, defaulting to
"pkg,goos,goarch".

Change-Id: I00413ab348bbff31743b59e81d88c4faab1a8dca
Reviewed-on: https://go-review.googlesource.com/38584
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

Change-Id: I8c284c7d8f1a567c740bfe78c2b8f1ee5e662e72
Reviewed-on: https://go-review.googlesource.com/38395Reviewed-by: Chris Broadfoot <cbro@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

Flush strings to disk before diffing files.

Change-Id: I91c47bfe352ee833a164fa7de00341e96498bc96
Reviewed-on: https://go-review.googlesource.com/39235Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

The Uploads table gets an entry when a user or the server aborts an
upload in mid-upload. The Records table is cleaned up by the
transaction, but the Uploads entry is left behind.

The optimized query that is used for the perf.golang.org homepage did
not properly exclude these uploads that have no records.

Change-Id: I8322af5946c233e5cb4b15bf57d9487bcd1e6423
Reviewed-on: https://go-review.googlesource.com/39051
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

Change-Id: I79f486a875e4c6858482d406b1c0328655e32aef
Reviewed-on: https://go-review.googlesource.com/38305
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

Change-Id: Ie2cafb6ee0fe30860803cc332133fc1922857831
Reviewed-on: https://go-review.googlesource.com/38304
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

Change-Id: I5c1363125998acf7d28692aa94ddfbef148577c2
Reviewed-on: https://go-review.googlesource.com/38094Reviewed-by: Quentin Smith <quentin@golang.org>

If there's just one benchmark, the geomean will be the same as the
benchmark result. Omit it.

Change-Id: Id165fa127aa8615015511d6a1da6e1be48b85539
Reviewed-on: https://go-review.googlesource.com/38093Reviewed-by: Quentin Smith <quentin@golang.org>

This makes the geomean undefined (or zero, depending on who you ask).
Since these usually come from counts, it's more informative to just
omit the benchmark.

Change-Id: Ia426955784cf12fea19b3d5e51a49a436cbc81c6
Reviewed-on: https://go-review.googlesource.com/38092Reviewed-by: Quentin Smith <quentin@golang.org>

Change-Id: Iceca4836257d102335be2cd4e86fae8fb869f652
Reviewed-on: https://go-review.googlesource.com/37853
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

- document local servers in README
- rename local servers to "localperf" and "localperfdata" so one can
  "go get" them.
- find static/templates dirs automatically
- allow configuring a persistent database in localperfdata

Change-Id: I4e62f23c38be6978f091ccbbda8002d9f588b8a4
Reviewed-on: https://go-review.googlesource.com/37717Reviewed-by: Russ Cox <rsc@golang.org>

This adds a -view_url_base flag to set the view URL base, and renames
-port to -addr to match analysis/localserver.

Change-Id: I2903ec0814c630345c212dac10a0f36567aab5ec
Reviewed-on: https://go-review.googlesource.com/37533Reviewed-by: Russ Cox <rsc@golang.org>

This changes benchsave to behave like benchstat; in addition to
uploading the files and printing a URL, it also prints the text format
of benchstat. This is fetched from the ViewURL provided by the storage
server, so the analysis can be changed/improved without requiring
users to rebuild benchsave.

Change-Id: I28519a5e3cf89962bd952ff26a8a6a717b9ef636
Reviewed-on: https://go-review.googlesource.com/37532Reviewed-by: Russ Cox <rsc@golang.org>

This uses a different query structure for ListUploads when the query
string is empty. On our production database, this reduces the query
time from ~5s to ~0.1s, which drastically improves the performance of
perf.golang.org.

Change-Id: I1f2bf97449858d1b06791dd352cc437dcbda1fac
Reviewed-on: https://go-review.googlesource.com/37168
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

SortBy seems to sometimes sort by the second column instead of the
first. Until that bug is figured out, do the sorting one column at a
time to control the ordering.

Without this patch, try
/trend?q=upload%3A20170216.3+name%3ARegexpMatchEasy0_1K&raw=1
and
/trend?q=trend%3Abuild.golang.org%2Fwindows-amd64-gce+name%3AGzip+commit-time%3E2016-11-08+commit-time%3C2016-11-30&raw=1

The former seems to sort first by "commit" instead of "commit-time",
but the latter sorts by "commit-time" first. Apply the patch and both
are sorted by commit-time.

Change-Id: I8ebd2720fa1147fc1d468a6bab515a224537a98e
Reviewed-on: https://go-review.googlesource.com/37471
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

Since the upload ID is part of the primary key, we can avoid using the
RecordLabels table to find matching records. This takes the query
"upload:20170127.22 name:Gzip" from ~0.3s to ~0.04s on our production
database.

Change-Id: Id31624ec1b81f3a7f648c6597392718c56a10edd
Reviewed-on: https://go-review.googlesource.com/37236
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

Change-Id: I7bd1ed03127830d6530415d77263ee8b0155a2c4
Reviewed-on: https://go-review.googlesource.com/37230
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

Change-Id: I0aad9567cc8887922541145868fd927d8c4fb10b
Reviewed-on: https://go-review.googlesource.com/37167
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

This adds an undocumented "raw=1" query parameter to visualize raw
results without any averaging or filtering. Results are colored
according to the upload-part they came from.

Change-Id: I25590ace89dfe55380e28182c38aeb223e2e73b9
Reviewed-on: https://go-review.googlesource.com/37160
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

Change-Id: I405ff8d89593365b0e27925e7640bee35c723682
Reviewed-on: https://go-review.googlesource.com/37158
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

Queries with multiple parts referring to the same key are now
simplified before sending to the database. If the query can be proven
to return no results, the query is not sent to the database at all.

Change-Id: I2d307d09f463fb0e6e7bd9b9902115916e7ffffa
Reviewed-on: https://go-review.googlesource.com/36953
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

This is still not feature complete, so the page is not linked from
anywhere. It supports graphing benchmark results that match any
perf query. One line is shown per matching benchmark name. Demo
available at:

https://quentin.perf.golang.org/trend?q=upload:20170127.25

Big TODO tasks:
- improve display of multiple benchmarks (multiple graphs?
  click-to-select? additional user controls?)
- correlate benchmarks across uploads (currently all uploads are
  assumed to be from identical machines)
- optimize/cache queries (6 months of a single benchmark loads in 12s;
  6 months of all benchmarks times out at >60s)

Change-Id: I7d1f6073c4837ef63205a10dd4e18085fe7d00ac
Reviewed-on: https://go-review.googlesource.com/36612
Run-TryBot: Quentin Smith <quentin@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

Change-Id: I89a2ad20d0378077fb90cae45fcc8cab33c147ec
Reviewed-on: https://go-review.googlesource.com/36952
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

With this CL, the perf repo should now pass all tests on Windows.

Change-Id: I91fead1e065f22b54236ff96c6ace529dc3964a8
Reviewed-on: https://go-review.googlesource.com/36627
Run-TryBot: Quentin Smith <quentin@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

The query term "key>" can be used to search for records that have any
value for the key. This optimizes the query with a separate SQL
expression and adds tests for this behavior.

Change-Id: I5e2734734e0911a3ed12c87e7ada776ad309a90d
Reviewed-on: https://go-review.googlesource.com/36590Reviewed-by: Russ Cox <rsc@golang.org>

Change-Id: I268e3731b51b2c7cb327fc47913058490c439d6a
Reviewed-on: https://go-review.googlesource.com/36610
Run-TryBot: Quentin Smith <quentin@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

- storage/db/dbtest, storage/cmd/reindex: hide cloud SQL behind a build tag
  (run go test -tags cloud golang.org/x/perf/storage/db -cloud to run tests)
- analysis/appengine, storage/appengine: hide behind appengine build tag
  (automatically supplied by goapp)
- vendor: used govendor to import non-x non-appengine dependencies

Change-Id: Ia5590599e7191ab3427618204fab7d2cec907579
Reviewed-on: https://go-review.googlesource.com/36592
Run-TryBot: Quentin Smith <quentin@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>

Change-Id: I08e93f469558ca3db356cb03eba345bfe29e36f8
Reviewed-on: https://go-review.googlesource.com/36252Reviewed-by: Russ Cox <rsc@golang.org>

Change-Id: I7f8c384dc9dd47b2706392188685a106a1ed487a
Reviewed-on: https://go-review.googlesource.com/36251Reviewed-by: Russ Cox <rsc@golang.org>

Change-Id: I8b4b85a2c38deb53bd30a26447aef8c6ce1b49d3
Reviewed-on: https://go-review.googlesource.com/36117Reviewed-by: Russ Cox <rsc@golang.org>

Change-Id: Id9eecd9bc5bf7be4ffcdd5a120c7ccc4afe2fb17
Reviewed-on: https://go-review.googlesource.com/36115Reviewed-by: Russ Cox <rsc@golang.org>

In a test database containing 309,000 real benchmark records, given
the query

  select t1.UploadID, (select t2.Value from RecordLabels t2 where
  t2.UploadID = t1.UploadID AND Name = 'upload-time' limit 1) from
  Uploads t1;

used to obtain sample values from each upload, the PRIMARY KEY index
causes the query to fall from 1.49s to 0.03s.

The NOT NULL constraint is necessary to make the PRIMARY KEY effective.

Change-Id: I9e663766e2f99f1096952e4c20eecbf8faab577b
Reviewed-on: https://go-review.googlesource.com/36114Reviewed-by: Russ Cox <rsc@golang.org>