Thread: PATCH: pgbench / int64 instead of int for xact count

PATCH: pgbench / int64 instead of int for xact count

From
Tomas Vondra
Date:
Hi,

I've been running a few longer pgbench tests (~week), and I've run into
this:

transaction type: SELECT only
scaling factor: 1250
query mode: simple
number of clients: 32
number of threads: 4
duration: 605000 s
number of transactions actually processed: -1785047856
latency average: -10.846 ms
tps = -2950.492090 (including connections establishing)
tps = -2950.492325 (excluding connections establishing)

The instance was doing ~10k tps for a week, which caused overflow of the
int counter, used to track number of transactions. Hence the negative
values.

I think we've reached the time when hardeare capable of doing this is
pretty common (SSDs, ...), so I think it's time to switch the counter to
int64.

Tomas

Attachment

Re: PATCH: pgbench / int64 instead of int for xact count

From
Andres Freund
Date:
Hi,

On 2014-05-25 18:05:03 +0200, Tomas Vondra wrote:
> I've been running a few longer pgbench tests (~week), and I've run into
> this:

> number of transactions actually processed: -1785047856
> latency average: -10.846 ms
> tps = -2950.492090 (including connections establishing)
> tps = -2950.492325 (excluding connections establishing)
> 
> The instance was doing ~10k tps for a week, which caused overflow of the
> int counter, used to track number of transactions. Hence the negative
> values.

> I think we've reached the time when hardeare capable of doing this is
> pretty common (SSDs, ...), so I think it's time to switch the counter to
> int64.

Especially when it's perfectly possible to do 500k read only
transactions a second...

>          printf("number of transactions per client: %d\n", nxacts);
> -        printf("number of transactions actually processed: %d/%d\n",
> +        printf("number of transactions actually processed: %ld/%d\n",
>                 normal_xacts, nxacts * nclients);

That's not right though. On windows a long (indicated by the %l) is only
4 bytes wide. Check INT64_FORMAT. That's generated by configure/platform
template files and should always be correct.

Greetings,

Andres Freund

-- Andres Freund                       http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services



Re: PATCH: pgbench / int64 instead of int for xact count

From
Tomas Vondra
Date:
On 25.5.2014 19:05, Andres Freund wrote:
>>          printf("number of transactions per client: %d\n", nxacts);
>> -        printf("number of transactions actually processed: %d/%d\n",
>> +        printf("number of transactions actually processed: %ld/%d\n",
>>                 normal_xacts, nxacts * nclients);
>
> That's not right though. On windows a long (indicated by the %l) is only
> 4 bytes wide. Check INT64_FORMAT. That's generated by configure/platform
> template files and should always be correct.

Oh, right. v2 of the patch attached.

Tomas

Attachment

Re: PATCH: pgbench / int64 instead of int for xact count

From
Tom Lane
Date:
Tomas Vondra <tv@fuzzy.cz> writes:
> On 25.5.2014 19:05, Andres Freund wrote:
>> That's not right though. On windows a long (indicated by the %l) is only
>> 4 bytes wide. Check INT64_FORMAT. That's generated by configure/platform
>> template files and should always be correct.

> Oh, right. v2 of the patch attached.

Hm ... if normal_xacts can overflow an int, shouldn't we expect that
the product nxacts * nclients could?
        regards, tom lane



Re: PATCH: pgbench / int64 instead of int for xact count

From
Tomas Vondra
Date:
On 25.5.2014 20:32, Tom Lane wrote:
> Tomas Vondra <tv@fuzzy.cz> writes:
>> On 25.5.2014 19:05, Andres Freund wrote:
>>> That's not right though. On windows a long (indicated by the %l) is only
>>> 4 bytes wide. Check INT64_FORMAT. That's generated by configure/platform
>>> template files and should always be correct.
>
>> Oh, right. v2 of the patch attached.
>
> Hm ... if normal_xacts can overflow an int, shouldn't we expect that
> the product nxacts * nclients could?

Maybe, but I saw that as a separate thing, mostly unrelated to the
'hardware got so fast it can overflow' problem. Because nxacts is the
parameter we use to define duration, so if it din't overflow in the
past, it won't overflow today.

OTOH, thanks to hardware improvements we tend to use more clients /
higher number of transactions, so fixing this seems like a good idea.

Should we change the type of nxacts to int64 (thus allowing '-t' with
64-bit integers), or just the overflow in the printf call? I don't find
it very practical to use -t with values not fitting into 32-bits (the -T
seems better to do that), so I'm inclined to just fix the printf. Patch
v3 attached.

Tomas

Attachment

Re: PATCH: pgbench / int64 instead of int for xact count

From
Tom Lane
Date:
Tomas Vondra <tv@fuzzy.cz> writes:
> Should we change the type of nxacts to int64 (thus allowing '-t' with
> 64-bit integers), or just the overflow in the printf call? I don't find
> it very practical to use -t with values not fitting into 32-bits (the -T
> seems better to do that), so I'm inclined to just fix the printf.

I concur: I don't see people setting -t higher than INT_MAX, but a -t
times -c product exceeding INT_MAX seems more plausible.

> Patch v3 attached.

Looks good, pushed.
        regards, tom lane