Thread: Re: [PERFORM] pgbench to the MAXINT

Em 10-01-2011 05:25, Greg Smith escreveu:
> Euler Taveira de Oliveira wrote:
>> Em 07-01-2011 22:59, Greg Smith escreveu:
>>> setrandom: invalid maximum number -2147467296
>>>
>> It is failing at atoi() circa pgbench.c:1036. But it just the first
>> one. There are some variables and constants that need to be converted
>> to int64 and some functions that must speak 64-bit such as getrand().
>> Are you working on a patch?
>
> http://archives.postgresql.org/pgsql-hackers/2010-01/msg02868.php
> http://archives.postgresql.org/message-id/4C326F46.4050801@2ndquadrant.com
>
Greg, I just improved your patch. I tried to work around the problems pointed
out in the above threads. Also, I want to raise some points:

(i) If we want to support and scale factor greater than 21474 we have to
convert some columns to bigint; it will change the test. From the portability
point it is a pity but as we have never supported it I'm not too worried about
it. Why? Because it will use bigint columns only if the scale factor is
greater than 21474. Is it a problem? I don't think so because generally people
compare tests with the same scale factor.

(ii) From the performance perspective, we need to test if the modifications
don't impact performance. I don't create another code path for 64-bit
modifications (it is too ugly) and I'm afraid some modifications affect the
32-bit performance. I'm in a position to test it though because I don't have a
big machine ATM. Greg, could you lead these tests?

(iii) I decided to copy scanint8() (called strtoint64 there) from backend
(Robert suggestion [1]) because Tom pointed out that strtoll() has portability
issues. I replaced atoi() with strtoint64() but didn't do any performance tests.

Comments?


[1] http://archives.postgresql.org/pgsql-hackers/2010-07/msg00173.php


--
   Euler Taveira de Oliveira
   http://www.timbira.com/

Euler Taveira de Oliveira wrote:
> (i) If we want to support and scale factor greater than 21474 we have 
> to convert some columns to bigint; it will change the test. From the 
> portability point it is a pity but as we have never supported it I'm 
> not too worried about it. Why? Because it will use bigint columns only 
> if the scale factor is greater than 21474. Is it a problem? I don't 
> think so because generally people compare tests with the same scale 
> factor.
>
> (ii) From the performance perspective, we need to test if the 
> modifications don't impact performance. I don't create another code 
> path for 64-bit modifications (it is too ugly) and I'm afraid some 
> modifications affect the 32-bit performance. I'm in a position to test 
> it though because I don't have a big machine ATM. Greg, could you lead 
> these tests?
>
> (iii) I decided to copy scanint8() (called strtoint64 there) from 
> backend (Robert suggestion [1]) because Tom pointed out that strtoll() 
> has portability issues. I replaced atoi() with strtoint64() but didn't 
> do any performance tests.

(i):  Completely agreed.

(ii):  There is no such thing as a "big machine" that is 32 bits now; 
anything that's 32 is a tiny system here in 2011.  What I can do is 
check for degredation on the only 32-bit system I have left here, my 
laptop.  I'll pick a sensitive test case and take a look.

(iii) This is an important thing to test, particularly given it has the 
potential to impact 64-bit results too.

Thanks for picking this up again and finishing the thing off.  I'll add 
this into my queue of performance tests to run and we can see if this is 
worth applying.  Probably take a little longer than the usual CF review 
time.  But as this doesn't interfere with other code people are working 
on and is sort of a bug fix, I don't think it will be a problem if it 
takes a little longer to get this done.

-- 
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us
"PostgreSQL 9.0 High Performance": http://www.2ndQuadrant.com/books

On Tue, Jan 18, 2011 at 1:42 PM, Greg Smith <greg@2ndquadrant.com> wrote:
> Thanks for picking this up again and finishing the thing off.  I'll add this
> into my queue of performance tests to run and we can see if this is worth
> applying.  Probably take a little longer than the usual CF review time.  But
> as this doesn't interfere with other code people are working on and is sort
> of a bug fix, I don't think it will be a problem if it takes a little longer
> to get this done.

At least in my book, we need to get this committed in the next two
weeks, or wait for 9.2.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Robert Haas wrote:
> At least in my book, we need to get this committed in the next two
> weeks, or wait for 9.2.
>   

Yes, I was just suggesting that I was not going to get started in the 
first week or two given the other pgbench related tests I had queued up 
already.  Those are closing up nicely, and I'll start testing 
performance of this change over the weekend.

-- 
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us
"PostgreSQL 9.0 High Performance": http://www.2ndQuadrant.com/books

The update on the work to push towards a bigger pgbench is that I now 
have the patch running and generating databases larger than any 
previously possible scale:

$ time pgbench -i -s 25000 pgbench
...
2500000000 tuples done.
...
real    258m46.350s
user    14m41.970s
sys    0m21.310s

$ psql -d pgbench -c "select 
pg_size_pretty(pg_relation_size('pgbench_accounts'));"pg_size_pretty
----------------313 GB

$ psql -d pgbench -c "select 
pg_size_pretty(pg_relation_size('pgbench_accounts_pkey'));"pg_size_pretty
----------------52 GB

$ time psql -d pgbench -c "select count(*) from pgbench_accounts"  count   
------------2500000000

real    18m48.363s
user    0m0.010s
sys    0m0.000s

The only thing wrong with the patch sent already needed to reach this 
point was this line:
    for (k = 0; k < naccounts * scale; k++)

Which needed a (int64) cast for the multiplied value in the middle there.

Unfortunately the actual test itself doesn't run yet.  Every line I see 
when running the SELECT-only test says:

client 0 sending SELECT abalance FROM pgbench_accounts WHERE aid = 1;

So something about the updated random generation code isn't quite right 
yet.  Now that I have this monster built, I'm going to leave it on the 
server until I can sort that out, which hopefully will finish up in the 
next day or so.

-- 
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us
"PostgreSQL 9.0 High Performance": http://www.2ndQuadrant.com/books

Attached is an updated 64-bit pgbench patch that works as expected for
all of the most common pgbench operations, including support for scales
above the previous boundary of just over 21,000.  Here's the patched
version running against a 303GB database with a previously unavailable
scale factor:

$ pgbench -T 300 -j 2 -c 4 pgbench
starting vacuum...end.
transaction type: TPC-B (sort of)
scaling factor: 25000
query mode: simple
number of clients: 4
number of threads: 2
duration: 300 s
number of transactions actually processed: 21681
tps = 72.249999 (including connections establishing)
tps = 72.250610 (excluding connections establishing)

And some basic Q/A that the values it touched were in the right range:

$ psql -d pgbench -c "select min(aid),max(aid) from pgbench_accounts";

 min |    max
-----+------------
   1 | 2500000000

$ psql -d pgbench -c "select min(aid),max(aid),count(*) from
pgbench_accounts where abalance!=0" &

  min  |    max     | count
-------+------------+-------
 51091 | 2499989587 | 21678

(This system was doing 300MB/s on reads while executing that count, and
it still took 19 minutes)

The clever way Euler updated the patch, you don't pay for the larger
on-disk data (bigint columns) unless you use a range that requires it,
which greatly reduces the number of ways the test results can suffer
from this change.  I felt the way that was coded was a bit more
complicated than it needed to be though, as it made where that switch
happened at get computed at runtime based on the true size of the
integers.  I took that complexity out and just put a hard line in there
instead:  if scale>=20000, you get bigints.  That's not very different
from the real limit, and it made documenting when the switch happens
easy to write and to remember.

The main performance concern with this change was whether using int64
more internally for computations would slow things down on a 32-bit
system.  I thought I'd test that on my few years old laptop.  It turns
out that even though I've been running an i386 Linux on here, it's
actually a 64-bit CPU.  (I think that it has a 32-bit install may be an
artifact of Adobe Flash install issues, sadly)  So this may not be as
good of a test case as I'd hoped.  Regardless, running a test aimed to
stress simple SELECTs, the thing I'd expect to suffer most from
additional CPU overhead, didn't show any difference in performance:

$ createdb pgbench
$ pgbench -i -s 10 pgbench
$ psql -c "show shared_buffers"
 shared_buffers
----------------
 256MB
(1 row)
$ pgbench -S -j 2 -c 4 -T 60 pgbench

i386    x86_64
6932    6924
6923    6926
6923    6922
6688    6772
6914    6791
6902    6916
6917    6909
6943    6837
6689    6744

6688    6744    min
6943    6926    max
6870    6860    average

Given the noise level of pgbench tests, I'm happy saying that is the
same speed.  I suspect the real overhead in pgbench's processing relates
to how it is constantly parsing text to turn them into statements, and
that how big the integers it uses are is barley detectable over that.

So...where does that leave this patch?  I feel that pgbench will become
less relevant very quickly in 9.1 unless something like this is
committed.  And there don't seem to be significant downsides to this in
terms of performance.  There are however a few rough points left in here
that might raise concern:

1) A look into the expected range of the rand() function suggests the
glibc implementation normally proves 30 bits of resolution, so about 1
billion numbers.  You'll have >1B rows in a pgbench database once the
scale goes over 10,000.  So without a major overhaul of how random
number generation is treated here, people can expect the distribution of
rows touched by a test run to get less even once the database scale gets
very large.  I added another warning paragraph to the end of the docs in
this update to mention this.  Long-term, I suspect we may need to adopt
a superior 64-bit RNG approach, something like a Mersenne Twister
perhaps.  That's a bit more than can be chewed on during 9.1 development
though.

2) I'd rate odds are good there's one or more corner-case bugs in
\setrandom or \setshell I haven't found yet, just from the way that code
was converted.  Those have some changes I haven't specifically tested
exhaustively yet.  I don't see any issues when running the most common
two pgbench tests, but that's doesn't mean every part of that 32 -> 64
bit conversion was done correctly.

Given how I use pgbench, for data generation and rough load testing, I'd
say neither of those concerns outweights the need to expand the size
range of this program.  I would be happy to see this go in, followed by
some alpha and beta testing aimed to see if any of the rough spots I'm
concerned about actually appear.  Unfortunately I can't fit all of those
tests in right now, as throwing around one of these 300GB data sets is
painful--when you're only getting 72 TPS, looking for large scale
patterns in the transactions takes a long time to do.  For example, if I
really wanted a good read on how bad the data distribution skew due to
small random range is, I'd need to let some things run for a week just
for a first pass.

I'd like to see this go in, but the problems I've spotted are such that
I would completely understand this being considered not ready by
others.  Just having this patch available here is a very useful step
forward in my mind, because now people can always just grab it and do a
custom build if they run into a larger system.

Wavering between Returned with Feedback and Ready for Committer here.
Thoughts?

--
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us
"PostgreSQL 9.0 High Performance": http://www.2ndQuadrant.com/books

diff --git a/contrib/pgbench/pgbench.c b/contrib/pgbench/pgbench.c
index 55ca1e8..5b9b582 100644
--- a/contrib/pgbench/pgbench.c
+++ b/contrib/pgbench/pgbench.c
@@ -60,6 +60,8 @@
 #define INT64_MAX    INT64CONST(0x7FFFFFFFFFFFFFFF)
 #endif

+#define MAX_RANDOM_VALUE64    INT64_MAX
+
 /*
  * Multi-platform pthread implementations
  */
@@ -364,15 +366,84 @@ usage(const char *progname)
            progname, progname);
 }

+/*
+ * strtoint64 -- convert a string to 64-bit integer
+ *
+ * This function is a modified version of scanint8() from
+ * src/backend/utils/adt/int8.c.
+ *
+ */
+static int64
+strtoint64(const char *str)
+{
+    const char *ptr = str;
+    int64        result = 0;
+    int            sign = 1;
+
+    /*
+     * Do our own scan, rather than relying on sscanf which might be broken
+     * for long long.
+     */
+
+    /* skip leading spaces */
+    while (*ptr && isspace((unsigned char) *ptr))
+        ptr++;
+
+    /* handle sign */
+    if (*ptr == '-')
+    {
+        ptr++;
+
+        /*
+         * Do an explicit check for INT64_MIN.    Ugly though this is, it's
+         * cleaner than trying to get the loop below to handle it portably.
+         */
+        if (strncmp(ptr, "9223372036854775808", 19) == 0)
+        {
+            result = -INT64CONST(0x7fffffffffffffff) - 1;
+            ptr += 19;
+            goto gotdigits;
+        }
+        sign = -1;
+    }
+    else if (*ptr == '+')
+        ptr++;
+
+    /* require at least one digit */
+    if (!isdigit((unsigned char) *ptr))
+        fprintf(stderr, "invalid input syntax for integer: \"%s\"\n", str);
+
+    /* process digits */
+    while (*ptr && isdigit((unsigned char) *ptr))
+    {
+        int64        tmp = result * 10 + (*ptr++ - '0');
+
+        if ((tmp / 10) != result)        /* overflow? */
+            fprintf(stderr, "value \"%s\" is out of range for type bigint\n", str);
+        result = tmp;
+    }
+
+gotdigits:
+
+    /* allow trailing whitespace, but not other trailing chars */
+    while (*ptr != '\0' && isspace((unsigned char) *ptr))
+        ptr++;
+
+    if (*ptr != '\0')
+        fprintf(stderr, "invalid input syntax for integer: \"%s\"\n", str);
+
+    return ((sign < 0) ? -result : result);
+}
+
 /* random number generator: uniform distribution from min to max inclusive */
-static int
-getrand(int min, int max)
+static int64
+getrand(int64 min, int64 max)
 {
     /*
      * Odd coding is so that min and max have approximately the same chance of
      * being selected as do numbers between them.
      */
-    return min + (int) (((max - min + 1) * (double) random()) / (MAX_RANDOM_VALUE + 1.0));
+    return min + (int64) (((max - min + 1) * (double) random()) / (MAX_RANDOM_VALUE + 1.0));
 }

 /* call PQexec() and exit() on failure */
@@ -887,7 +958,7 @@ top:
         if (commands[st->state] == NULL)
         {
             st->state = 0;
-            st->use_file = getrand(0, num_files - 1);
+            st->use_file = (int) getrand(0, num_files - 1);
             commands = sql_files[st->use_file];
         }
     }
@@ -1007,7 +1078,7 @@ top:
         if (pg_strcasecmp(argv[0], "setrandom") == 0)
         {
             char       *var;
-            int            min,
+            int64        min,
                         max;
             char        res[64];

@@ -1019,15 +1090,15 @@ top:
                     st->ecnt++;
                     return true;
                 }
-                min = atoi(var);
+                min = strtoint64(var);
             }
             else
-                min = atoi(argv[2]);
+                min = strtoint64(argv[2]);

 #ifdef NOT_USED
             if (min < 0)
             {
-                fprintf(stderr, "%s: invalid minimum number %d\n", argv[0], min);
+                fprintf(stderr, "%s: invalid minimum number " INT64_FORMAT "\n", argv[0], min);
                 st->ecnt++;
                 return;
             }
@@ -1041,22 +1112,22 @@ top:
                     st->ecnt++;
                     return true;
                 }
-                max = atoi(var);
+                max = strtoint64(var);
             }
             else
-                max = atoi(argv[3]);
+                max = strtoint64(argv[3]);

-            if (max < min || max > MAX_RANDOM_VALUE)
+            if (max < min || max > MAX_RANDOM_VALUE64)
             {
-                fprintf(stderr, "%s: invalid maximum number %d\n", argv[0], max);
+                fprintf(stderr, "%s: invalid maximum number " INT64_FORMAT "\n", argv[0], max);
                 st->ecnt++;
                 return true;
             }

 #ifdef DEBUG
-            printf("min: %d max: %d random: %d\n", min, max, getrand(min, max));
+            printf("min: " INT64_FORMAT " max: " INT64_FORMAT " random: " INT64_FORMAT "\n", min, max, getrand(min,
max));
 #endif
-            snprintf(res, sizeof(res), "%d", getrand(min, max));
+            snprintf(res, sizeof(res), INT64_FORMAT, getrand(min, max));

             if (!putVariable(st, argv[0], argv[1], res))
             {
@@ -1069,7 +1140,7 @@ top:
         else if (pg_strcasecmp(argv[0], "set") == 0)
         {
             char       *var;
-            int            ope1,
+            int64        ope1,
                         ope2;
             char        res[64];

@@ -1081,13 +1152,13 @@ top:
                     st->ecnt++;
                     return true;
                 }
-                ope1 = atoi(var);
+                ope1 = strtoint64(var);
             }
             else
-                ope1 = atoi(argv[2]);
+                ope1 = strtoint64(argv[2]);

             if (argc < 5)
-                snprintf(res, sizeof(res), "%d", ope1);
+                snprintf(res, sizeof(res), INT64_FORMAT, ope1);
             else
             {
                 if (*argv[4] == ':')
@@ -1098,17 +1169,17 @@ top:
                         st->ecnt++;
                         return true;
                     }
-                    ope2 = atoi(var);
+                    ope2 = strtoint64(var);
                 }
                 else
-                    ope2 = atoi(argv[4]);
+                    ope2 = strtoint64(argv[4]);

                 if (strcmp(argv[3], "+") == 0)
-                    snprintf(res, sizeof(res), "%d", ope1 + ope2);
+                    snprintf(res, sizeof(res), INT64_FORMAT, ope1 + ope2);
                 else if (strcmp(argv[3], "-") == 0)
-                    snprintf(res, sizeof(res), "%d", ope1 - ope2);
+                    snprintf(res, sizeof(res), INT64_FORMAT, ope1 - ope2);
                 else if (strcmp(argv[3], "*") == 0)
-                    snprintf(res, sizeof(res), "%d", ope1 * ope2);
+                    snprintf(res, sizeof(res), INT64_FORMAT, ope1 * ope2);
                 else if (strcmp(argv[3], "/") == 0)
                 {
                     if (ope2 == 0)
@@ -1117,7 +1188,7 @@ top:
                         st->ecnt++;
                         return true;
                     }
-                    snprintf(res, sizeof(res), "%d", ope1 / ope2);
+                    snprintf(res, sizeof(res), INT64_FORMAT, ope1 / ope2);
                 }
                 else
                 {
@@ -1239,9 +1310,9 @@ init(void)
         "drop table if exists pgbench_tellers",
         "create table pgbench_tellers(tid int not null,bid int,tbalance int,filler char(84)) with (fillfactor=%d)",
         "drop table if exists pgbench_accounts",
-        "create table pgbench_accounts(aid int not null,bid int,abalance int,filler char(84)) with (fillfactor=%d)",
+        "create table pgbench_accounts(aid %s not null,bid int,abalance int,filler char(84)) with (fillfactor=%d)",
         "drop table if exists pgbench_history",
-        "create table pgbench_history(tid int,bid int,aid int,delta int,mtime timestamp,filler char(22))"
+        "create table pgbench_history(tid int,bid int,aid %s,delta int,mtime timestamp,filler char(22))"
     };
     static char *DDLAFTERs[] = {
         "alter table pgbench_branches add primary key (bid)",
@@ -1253,6 +1324,7 @@ init(void)
     PGresult   *res;
     char        sql[256];
     int            i;
+    int64        k;

     if ((con = doConnect()) == NULL)
         exit(1);
@@ -1263,8 +1335,7 @@ init(void)
          * set fillfactor for branches, tellers and accounts tables
          */
         if ((strstr(DDLs[i], "create table pgbench_branches") == DDLs[i]) ||
-            (strstr(DDLs[i], "create table pgbench_tellers") == DDLs[i]) ||
-            (strstr(DDLs[i], "create table pgbench_accounts") == DDLs[i]))
+            (strstr(DDLs[i], "create table pgbench_tellers") == DDLs[i]))
         {
             char        ddl_stmt[128];

@@ -1272,6 +1343,36 @@ init(void)
             executeStatement(con, ddl_stmt);
             continue;
         }
+        else if (strstr(DDLs[i], "create table pgbench_accounts") == DDLs[i])
+        {
+            char        ddl_stmt[128];
+
+            /*
+             * Use bigint columns in cases where scale factor is bigger
+             * than 20000, which gives an account number around the upper
+             * limit the range for a regular integer (just over 2 billion).
+             * For smaller scale factors, we still use int columns, to keep
+             * results more like those given by earlier versions of pgbench.
+             */
+            if (scale >= 20000)
+                snprintf(ddl_stmt, 128, DDLs[i], "bigint", fillfactor);
+            else
+                snprintf(ddl_stmt, 128, DDLs[i], "int", fillfactor);
+            executeStatement(con, ddl_stmt);
+            continue;
+        }
+        else if (strstr(DDLs[i], "create table pgbench_history") == DDLs[i])
+        {
+            char        ddl_stmt[128];
+
+            /* The accounts value in the history table needs a matching size */
+            if (scale >= 20000)
+                snprintf(ddl_stmt, 128, DDLs[i], "bigint");
+            else
+                snprintf(ddl_stmt, 128, DDLs[i], "int");
+            executeStatement(con, ddl_stmt);
+            continue;
+        }
         else
             executeStatement(con, DDLs[i]);
     }
@@ -1309,11 +1410,11 @@ init(void)
     }
     PQclear(res);

-    for (i = 0; i < naccounts * scale; i++)
+    for (k = 0; k < (int64) naccounts * scale; k++)
     {
-        int            j = i + 1;
+        int64            j = k + 1;

-        snprintf(sql, 256, "%d\t%d\t%d\t\n", j, i / naccounts + 1, 0);
+        snprintf(sql, 256, INT64_FORMAT "\t" INT64_FORMAT "\t%d\t\n", j, k / naccounts + 1, 0);
         if (PQputline(con, sql))
         {
             fprintf(stderr, "PQputline failed\n");
@@ -1321,7 +1422,7 @@ init(void)
         }

         if (j % 10000 == 0)
-            fprintf(stderr, "%d tuples done.\n", j);
+            fprintf(stderr, INT64_FORMAT " tuples done.\n", j);
     }
     if (PQputline(con, "\\.\n"))
     {
diff --git a/doc/src/sgml/pgbench.sgml b/doc/src/sgml/pgbench.sgml
index a33ac17..05c9794 100644
--- a/doc/src/sgml/pgbench.sgml
+++ b/doc/src/sgml/pgbench.sgml
@@ -155,6 +155,10 @@ pgbench <optional> <replaceable>options</> </optional> <replaceable>dbname</>
         Multiply the number of rows generated by the scale factor.
         For example, <literal>-s 100</> will create 10,000,000 rows
         in the <structname>pgbench_accounts</> table. Default is 1.
+        When the scale is 20,000 or larger, the columns used to
+        hold account identifiers will switch to using larger integers,
+        in order to be big enough to hold the range of account
+        identifiers.
        </para>
       </listitem>
      </varlistentry>
@@ -769,6 +773,19 @@ statement latencies in milliseconds:
    instances concurrently, on several client machines, against the same
    database server.
   </para>
+
+  <para>
+   The random number generation done by <application>pgbench</> is limited
+   the accuracy of the underlying operating system numeric library.  Typically
+   this provides range of approximately 1 billion values.  There will
+   be a billion rows in the <structname>pgbench_accounts</> table once the
+   database scale is increased to 10,000 or above.  When creating large
+   databases above or even near this scale, you should expect that the
+   distribution of rows selected by random operations will no longer be
+   as evenly distributed, relative to when the program is running against a
+   smaller database.
+  </para>
+
  </sect2>

 </sect1>

Greg,

* Greg Smith (greg@2ndquadrant.com) wrote:
> I took that complexity out and just put a hard line
> in there instead:  if scale>=20000, you get bigints.  That's not
> very different from the real limit, and it made documenting when the
> switch happens easy to write and to remember.

Agreed completely on this.

> It turns out that even though I've been running an i386 Linux on
> here, it's actually a 64-bit CPU.  (I think that it has a 32-bit
> install may be an artifact of Adobe Flash install issues, sadly)  So
> this may not be as good of a test case as I'd hoped.

Actually, I would think it'd still be sufficient..  If you're under a
32bit kernel you're not going to be using the extended registers, etc,
that would be available under a 64bit kernel..  That said, the idea that
we should care about 32-bit systems these days, in a benchmarking tool,
is, well, silly, imv.

> 1) A look into the expected range of the rand() function suggests
> the glibc implementation normally proves 30 bits of resolution, so
> about 1 billion numbers.  You'll have >1B rows in a pgbench database
> once the scale goes over 10,000.  So without a major overhaul of how
> random number generation is treated here, people can expect the
> distribution of rows touched by a test run to get less even once the
> database scale gets very large.

Just wondering, did you consider just calling random() twice and
smashing the result together..?

> I added another warning paragraph
> to the end of the docs in this update to mention this.  Long-term, I
> suspect we may need to adopt a superior 64-bit RNG approach,
> something like a Mersenne Twister perhaps.  That's a bit more than
> can be chewed on during 9.1 development though.

I tend to agree that we should be able to improve the random number
generation in the future.  Additionally, imv, we should be able to say
"pg_bench version X isn't comparable to version Y" in the release notes
or something, or have seperate version #s for it which make it clear
what can be compared to each other and what can't.  Painting ourselves
into a corner by saying we can't ever make pgbench generate results that
can't be compared to every other released version of pgbench just isn't
practical.

> 2) I'd rate odds are good there's one or more corner-case bugs in
> \setrandom or \setshell I haven't found yet, just from the way that
> code was converted.  Those have some changes I haven't specifically
> tested exhaustively yet.  I don't see any issues when running the
> most common two pgbench tests, but that's doesn't mean every part of
> that 32 -> 64 bit conversion was done correctly.

I'll take a look. :)
Thanks,
    Stephen

Stephen Frost wrote:
> Just wondering, did you consider just calling random() twice and
> smashing the result together..?
>   

I did.  The problem is that even within the 32 bits that random() 
returns, it's not uniformly distributed.  Combining two of them isn't 
really going to solve the distribution problem, just move it around.  
Some number of lower-order bits are less random than the others, and 
which they are is implementation dependent.

Poking around a bit more, I just discovered another possible approach is 
to use erand48 instead of rand in pgbench, which is either provided by 
the OS or emulated in src/port/erand48.c  That's way more resolution 
than needed here, given that 2^48 pgbench accounts would be a scale of 
2.8M, which makes for a database of about 42 petabytes.

-- 
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us
"PostgreSQL 9.0 High Performance": http://www.2ndQuadrant.com/books

Greg Smith <greg@2ndquadrant.com> writes:
> Poking around a bit more, I just discovered another possible approach is 
> to use erand48 instead of rand in pgbench, which is either provided by 
> the OS or emulated in src/port/erand48.c  That's way more resolution 
> than needed here, given that 2^48 pgbench accounts would be a scale of 
> 2.8M, which makes for a database of about 42 petabytes.

I think that might be a good idea --- it'd reduce the cross-platform
variability of the results quite a bit, I suspect.  random() is not
to be trusted everywhere, but I think erand48 is pretty much the same
wherever it exists at all (and src/port/ provides it elsewhere).
        regards, tom lane

Greg,

* Tom Lane (tgl@sss.pgh.pa.us) wrote:
> Greg Smith <greg@2ndquadrant.com> writes:
> > Poking around a bit more, I just discovered another possible approach is
> > to use erand48 instead of rand in pgbench, which is either provided by
> > the OS or emulated in src/port/erand48.c  That's way more resolution
> > than needed here, given that 2^48 pgbench accounts would be a scale of
> > 2.8M, which makes for a database of about 42 petabytes.
>
> I think that might be a good idea --- it'd reduce the cross-platform
> variability of the results quite a bit, I suspect.  random() is not
> to be trusted everywhere, but I think erand48 is pretty much the same
> wherever it exists at all (and src/port/ provides it elsewhere).

Works for me.  Greg, will you be able to work on this change?  If not, I
might be able to.
Thanks,
    Stephen

On Fri, Feb 11, 2011 at 8:35 AM, Stephen Frost <sfrost@snowman.net> wrote:
> Greg,
>
> * Tom Lane (tgl@sss.pgh.pa.us) wrote:
>> Greg Smith <greg@2ndquadrant.com> writes:
>> > Poking around a bit more, I just discovered another possible approach is
>> > to use erand48 instead of rand in pgbench, which is either provided by
>> > the OS or emulated in src/port/erand48.c  That's way more resolution
>> > than needed here, given that 2^48 pgbench accounts would be a scale of
>> > 2.8M, which makes for a database of about 42 petabytes.
>>
>> I think that might be a good idea --- it'd reduce the cross-platform
>> variability of the results quite a bit, I suspect.  random() is not
>> to be trusted everywhere, but I think erand48 is pretty much the same
>> wherever it exists at all (and src/port/ provides it elsewhere).
>
> Works for me.  Greg, will you be able to work on this change?  If not, I
> might be able to.

Seeing as how this patch has not been updated, I think it's time to
mark this one Returned with Feedback.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Tom Lane wrote:
> I think that might be a good idea --- it'd reduce the cross-platform
> variability of the results quite a bit, I suspect.  random() is not
> to be trusted everywhere, but I think erand48 is pretty much the same
> wherever it exists at all (and src/port/ provides it elsewhere).
>   

Given that pgbench will run with threads in some multi-worker 
configurations, after some more portability research I think odds are 
good we'd get nailed by 
http://sourceware.org/bugzilla/show_bug.cgi?id=10320 : "erand48 
implementation not thread safe but POSIX says it should be".  The AIX 
docs have a similar warning on them, so who knows how many versions of 
that library have the same issue.

Maybe we could make sure the one in src/port/ is thread safe and make 
sure pgbench only uses it.  This whole area continues to be messy enough 
that I think the patch needs to brew for another CF before it will all 
be sorted out properly.  I'll mark it accordingly and can pick this back 
up later.

-- 
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD

Greg Smith <greg@2ndquadrant.com> writes:
> Given that pgbench will run with threads in some multi-worker 
> configurations, after some more portability research I think odds are 
> good we'd get nailed by 
> http://sourceware.org/bugzilla/show_bug.cgi?id=10320 : "erand48 
> implementation not thread safe but POSIX says it should be".  The AIX 
> docs have a similar warning on them, so who knows how many versions of 
> that library have the same issue.

FWIW, I think that bug report is effectively complaining that if you use
both drand48 and erand48, the former can impact the latter.  If you use
only erand48, I don't see that there's any problem.
        regards, tom lane

Hi,

I have reviewed this patch.

https://commitfest.postgresql.org/action/patch_view?id=1068

2012/12/21 Gurjeet Singh <singh.gurjeet@gmail.com>:
>     The patch is very much what you had posted, except for a couple of
> differences due to bit-rot. (i) I didn't have to #define MAX_RANDOM_VALUE64
> since its cousin MAX_RANDOM_VALUE is not used by code anymore, and (ii) I
> used ternary operator in DDLs[] array to decide when to use bigint vs int
> columns.
>
>     Please review.
>
>     As for tests, I am currently running 'pgbench -i -s 21474' using
> unpatched pgbench, and am recording the time taken;Scale factor 21475 had
> actually failed to do anything meaningful using unpatched pgbench. Next I'll
> run with '-s 21475' on patched version to see if it does the right thing,
> and in acceptable time compared to '-s 21474'.
>
>     What tests would you and others like to see, to get some confidence in
> the patch? The machine that I have access to has 62 GB RAM, 16-core
> 64-hw-threads, and about 900 GB of disk space.

I have tested this patch, and hvae confirmed that the columns
for aid would be switched to using bigint, instead of int,
when the scalefactor >= 20,000.
(aid columns would exeed the upper bound of int when sf>21474.)

Also, I added a few fixes on it.

- Fixed to apply for the current git master.
- Fixed to surpress few more warnings about INT64_FORMAT.
- Minor improvement in the docs. (just my suggestion)

I attached the revised one.

Regards,
--
Satoshi Nagayasu <snaga@uptime.jp>
Uptime Technologies, LLC http://www.uptime.jp/

On 28.01.2013 23:30, Gurjeet Singh wrote:
> On Sat, Jan 26, 2013 at 11:24 PM, Satoshi Nagayasu<snaga@uptime.jp>  wrote:
>
>> 2012/12/21 Gurjeet Singh<singh.gurjeet@gmail.com>:
>>>      The patch is very much what you had posted, except for a couple of
>>> differences due to bit-rot. (i) I didn't have to #define
>> MAX_RANDOM_VALUE64
>>> since its cousin MAX_RANDOM_VALUE is not used by code anymore, and (ii) I
>>> used ternary operator in DDLs[] array to decide when to use bigint vs int
>>> columns.
>>>
>>>      Please review.
>>>
>>>      As for tests, I am currently running 'pgbench -i -s 21474' using
>>> unpatched pgbench, and am recording the time taken;Scale factor 21475 had
>>> actually failed to do anything meaningful using unpatched pgbench. Next
>> I'll
>>> run with '-s 21475' on patched version to see if it does the right thing,
>>> and in acceptable time compared to '-s 21474'.
>>>
>>>      What tests would you and others like to see, to get some confidence
>> in
>>> the patch? The machine that I have access to has 62 GB RAM, 16-core
>>> 64-hw-threads, and about 900 GB of disk space.
>>
>> I have tested this patch, and hvae confirmed that the columns
>> for aid would be switched to using bigint, instead of int,
>> when the scalefactor>= 20,000.
>> (aid columns would exeed the upper bound of int when sf>21474.)
>>
>> Also, I added a few fixes on it.
>>
>> - Fixed to apply for the current git master.
>> - Fixed to surpress few more warnings about INT64_FORMAT.
>> - Minor improvement in the docs. (just my suggestion)
>>
>> I attached the revised one.
>
> Looks good to me. Thanks!

Ok, committed.

At some point, we might want to have a strtoll() implementation in src/port.

- Heikki