This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
Re: [RFC][patch] Make DCACHE_LINE runtime-settable
- From: Jan Kratochvil <jan dot kratochvil at redhat dot com>
- To: Paul Pluzhnikov <ppluzhnikov at google dot com>
- Cc: gdb-patches at sourceware dot org
- Date: Mon, 25 Jul 2011 21:00:08 +0200
- Subject: Re: [RFC][patch] Make DCACHE_LINE runtime-settable
- References: <20110722222025.ED9B6190B14@elbrus2.mtv.corp.google.com>
On Sat, 23 Jul 2011 00:20:25 +0200, Paul Pluzhnikov wrote:
> For a 200ms round-trip connection, executing "where" and getting back 30
> frames translates into 6 seconds of wall time, and "thread apply all where"
> into 1 to 10 to 100 minutes (depending on how many threads there are,
> with 100 threads being common and 1000 not being exceptional).
>
> Attached patch allows the user to e.g. "set dcache-line-size 4096" and
> drastically reduce the number of round trips.
I was tweaking the LINE_SIZE_POWER facing this problem in mail:
https://fedorahosted.org/pipermail/crash-catcher/2010-December/001441.html
I have seen some of the target_read_memory requests are needlessly fragmented
into LINE_SIZE_POWER sized read requests. Have you considered making the
gdbserver protocol read requests size dynamic depending on the caller's
requested read size?
After all I found it is in many cases the fastest to upload the whole core
file as then there are no RTT delays at all but that is offtopic here.
Thanks,
Jan