Re: [E] Proposed performance improvement of tagged debug output

[Susan Hinrichs <sh...@verizonmedia.com.INVALID>]

Sounds encouraging.  I would suggest going with option 2.  Otherwise,
very few people will bother learning the new, higher performance
interface.

On Mon, Feb 8, 2021 at 5:46 PM Walt Karas
<wk...@verizonmedia.com.invalid> wrote:
>
> https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_apache_trafficserver_pull_7452&d=DwIBaQ&c=sWW_bEwW_mLyN3Kx2v57Q8e-CRbmiT9yOhqES_g_wVY&r=GAqM_xZpxNbVqsR-aGvQBjOG3d33Y2-i4ynL-JkEouY&m=T3uKuRdXuJZJ3DyMoQBcYkbjgx3A0sYIMfodLs6GqEo&s=bHquMyBVeX-FU8jZM7jR_NnfBCUBghgDxLgfL9ZvzVU&e=
>
> This PR changes TS API.  It introduces this new type:
>
> typedef struct TSFDbgCtl_s {
>   char volatile on; // Flag
>   char const *tag;
> } TSFDbgCtl;
>
> And this function:
>
> tsapi TSFDbgCtl const *TSFDbgCtlCreate(char const *tag);
>
> And this new printf-like macro:
>
> #define TSFDbg(ctlp, ...)
>
> Where the first parameter is a pointer returned by TSFDbgCtlCreate(), the
> second a printf format string, and the rest the parameters referenced the
> format string.
>
> In my performance test, enabling the new type of debug output resulted in a
> 10% reduction in requests per second handled, versus a 95% reduction due to
> enabling the old type of debug output. (See
> https://urldefense.proofpoint.com/v2/url?u=https-3A__gist.github.com_ywkaras_4b0386341a4a6ce0afacaf46dcce79d6&d=DwIBaQ&c=sWW_bEwW_mLyN3Kx2v57Q8e-CRbmiT9yOhqES_g_wVY&r=GAqM_xZpxNbVqsR-aGvQBjOG3d33Y2-i4ynL-JkEouY&m=T3uKuRdXuJZJ3DyMoQBcYkbjgx3A0sYIMfodLs6GqEo&s=UqdDEzH8fQFsDoVAxnY0X6060A6Go_eBpyvilYn90LI&e=  )
>
> If we decide to proceed with this, I think we have two options:
>
>
>    1.  Proceed with the approach where enabling low-overhead debug output
>    will only work for Debug()/TSFDbg calls that use a DbgCtl/TSDbgCtl instance
>    rather than a simple tag string.  This would include adding doc changes and
>    Au testing.
>    2. Switch to an approach similar to what I used for testing.  Where
>    Debug() calls and is_debug_tag_set() will (each) generate a hidden instance
>    of DbgCtl.  I think I can also make TSDebug() generate a hidden pointer to
>    TSFDbgCt.