home
***
CD-ROM
|
disk
|
FTP
|
other
***
search
/
SGI Varsity Update 1998 August
/
SGI Varsity Update 1998 August.iso
/
relnotes
/
SpeedShop
/
ch5.z
/
ch5
Wrap
Text File
|
1998-07-29
|
8KB
|
199 lines
- 1 -
5. _K_n_o_w_n__P_r_o_b_l_e_m_s__a_n_d__W_o_r_k_a_r_o_u_n_d_s
+o MPI 3.1.x.x reorganizes the dynamic objects specified
by _RLD_LIST and breaks ssrun's interposition
(incidents #567877 and #599278). The workaround is to
set environment variable MPI_RLD_HACK_OFF.
+o Any SpeedShop experiment will not produce data if the
process is killed with a SIGKILL signal. Therefore, any
experiment run under the WorkShop debugger will not
produce data if the process is killed from the user
interface. Doing so sends a SIGKILL, which can not be
caught, so that the data is lost. The workaround is to
kill the process by sending it a ctrl-C signal
(SIGINT).
+o _s_s_r_u_n -_i_d_e_a_l will not produce an experiment if it is
run in a directory where a previous direct invocation
of _p_i_x_i_e was done on the same executable. It will not
add the SpeedShop runtime, but will produce a .Counts
file instead. Workaround is to remove the pixified
executable before invoking _s_s_r_u_n. [Incident 517224]
+o Executables built non-shared are not generally
supported (except for direct _p_i_x_i_e use).
+o Usertime experiments on o32 programs can not collect
callstacks unless the executable has been explicitly
linked with -_l_e_x_c. [Unfixable; a fix would increase
the size of all executables, and is considered
undesirable for that reason.] The symptom of such
experiments is that all time is attributable only to
the current PC, with no attribution to callers at all.
+o On o32 programs there is no accurate way of unwinding
the stack in the prologue or epilogue using the PDR's.
Usertime experiments might fail to correctly unwind the
stack if the program is interrupted inside the prologue
or epilogue of a function. [Incident 366232]
+o Usertime experiments might fail inside function
_u_s_i_n_i_t() in n32/n64 _l_i_b_c._s_o._1 on IRIX 6.2 because of
incorrect dwarf information. [Incident 439873]
+o Programs using _l_i_b_m_p._s_o and _l_i_b_f_t_n._s_o might not produce
complete experiment files if they are terminated with
SIGINT. As a workaround, environment variable
_SPEEDSHOP_SHUTDOWN_SIG can be used. [Incident 406405]
+o _p_r_o_f does not handle the -dis or -source options for
call stack trace experiments. [Incident 409155]
- 2 -
+o _p_r_o_f does not handle the -h or -l options for call
stack trace experiments. [Incident 531749]
+o _p_r_o_f on usertime experiments will overcount inclusive
time (and on fpe experiments might overcount inclusive
fpes) on programs that have recursive functions. Each
level of recursion will count for each callstack. This
may lead to functions whose inclusive time greatly
exceeds the actual runtime. [Incident 438769]
+o Butterfly display in _p_r_o_f needs to show more
descriptive information, including the callsite
addresses and the invocation counts. [Incident 527599]
+o _p_r_o_f doesn't always handle correctly hidden symbol
resolution between dso's when associating performance
numbers with those symbols, so there may be some
inaccuracies for those symbols that have multiple
definitions from different dso's. [Incident 562642]
+o _p_r_o_f has still some performance problems. [Incident
537714]
+o Some marching orders combinations in experiment files
cause _p_r_o_f to break. [Incident 578807]
+o The butterfly views produced by prof may be incorrect
for some symbols. [Incident 579434]
+o _p_r_o_f -feedback may core dump in some cases. [Incident
594034]
+o _p_r_o_f doesn't yet use targinfo to simulate the hardware
architecture when processing ideal experiment data, so
some cost attributions may be inaccurate. [Incident
505566]
+o Ideal experiments might not produce correct call counts
for procedures with multiple entry points and might
miss some intra-DSO calls. [Incident 260827]
+o Totaltime experiments on IRIX 6.2 or later and hwctime
experiments on IRIX 6.5 or later might produce biased
performance data for programs using pthreads, since
ITIMER_REAL in this case is asynchronous and SIGALRM
gets delivered to the process (and not to all the
uthreads) and he o/s hw counter access interface does
not provide per-uthread controls. [Incident 477551]
+o _f_p_c_s_a_m_p (1 msec sampling) experiments on R10000 systems
on IRIX 6.2 or 6.3 might produce inconsistent
- 3 -
performance results because of a kernel bug. [Incident
409775]
+o There is a potential problem with ideal experiments and
instrumentation on MPI programs running on different
hosts but under the same (NFS-mounted) directory. There
may turn out to be two _p_i_x_i_e executables running on two
different hosts, producing a.out.pixie into a single
nfs-mounted directory at the same time. For DSOs, the
problem is more complex, because there may be two
different libc's involved, one from each host, but
_p_i_x_i_e is producing a libc.so.1.pix32, and only one can
live in the directory. As a workaround, the user can
specify a different directory on each host using _m_p_i_r_u_n
-d. [Incident 506183]
+o Performance analysis using either prof or WorkShop's
performance analyzer can only happen on the host that
generated the experiment files. For kernel profiling
experiments there's a workaround described in this
section.
+o If a kernel profiling experiment has been run on a
remote system or on the same system but using a
different kernel file than "/unix", then the kernel
file name should be provided as a command-line argument
to prof (or, if WorkShop's performance analyzer is used
to process the experiment file, the kernel file should
be copied into the same directory as the experiment
file and should be renamed to "unix").
+o Usage and heartbeat data include resources used up by
instrumentation, so their values do not accurately
reflect the performance data of just the target
application for those experiments that require binary
code instrumentation. [Incident 577985]
+o SpeedShop doesn't support per-pthread performance data
for all experiments. Only usertime, totaltime, hwctime,
fpe, io and heap experiments tag collected data with
pthread ids for later per-pthread analysis. [Incident
476886]
+o PVM support is under investigation. [Incident 581107]
+o The 6.5 MR corded libSgm.so.1 fails when pixified.
[Incident 608528]