Add unscheduled-thread modeling to the drmemtrace scheduler #6822
Assignees
Labels
Comments
derekbruening
added a commit
that referenced
this issue
May 21, 2024
Refactors the scheduler marker processing into two virtual functions subclasses can override: the initialization time header marker reading now has process_next_initial_record() while regular processing has process_marker(). No change in functionality. Future work will add "unscheduled thread" support with new markers, building on this refactoring. Issue: #6822
derekbruening
added a commit
that referenced
this issue
May 21, 2024
Refactors the scheduler marker processing into two virtual functions subclasses can override: the initialization time header marker reading now has process_next_initial_record() while regular processing has process_marker(). No change in functionality. Future work will add "unscheduled thread" support with new markers, building on this refactoring. Issue: #6822
derekbruening
added a commit
that referenced
this issue
May 24, 2024
Augments the drmemtrace scheduler with a new notion of an "unscheduled" input which, if it has no timeout, is not runnable indefinitely until another input explicitly wakes it up. Adds 3 new markers: + TRACE_MARKER_TYPE_SYSCALL_UNSCHEDULE, which makes the caller "unscheduled". + TRACE_MARKER_TYPE_SYSCALL_SCHEDULE, which makes a target no longer "unscheduled". + TRACE_MARKER_TYPE_SYSCALL_ARG_TIMEOUT which adds a timeout parameter to TRACE_MARKER_TYPE_SYSCALL_UNSCHEDULE and TRACE_MARKER_TYPE_DIRECT_THREAD_SWITCH Adds handling for the new marker types. Changes TRACE_MARKER_TYPE_DIRECT_THREAD_SWITCH to make the source input unscheduled, unless a timeout is passed which is honored (previously it was ignored and the syscall latency was used as the block time if it was over the thresholds). Adds a fallback in case unscheduled inputs are left with no schedulable inputs. Adds unit tests. Also tested on a large trace with many real-world cases of these direct switches and unschedulable threads. Issue: #6822
derekbruening
added a commit
that referenced
this issue
May 29, 2024
Augments the drmemtrace scheduler with a new notion of an "unscheduled" input which, if it has no timeout, is not runnable indefinitely until another input explicitly wakes it up. Adds 3 new markers: + TRACE_MARKER_TYPE_SYSCALL_UNSCHEDULE, which makes the caller "unscheduled". + TRACE_MARKER_TYPE_SYSCALL_SCHEDULE, which makes a target no longer "unscheduled". + TRACE_MARKER_TYPE_SYSCALL_ARG_TIMEOUT which adds a timeout parameter to TRACE_MARKER_TYPE_SYSCALL_UNSCHEDULE and TRACE_MARKER_TYPE_DIRECT_THREAD_SWITCH Adds handling for the new marker types. Changes TRACE_MARKER_TYPE_DIRECT_THREAD_SWITCH to make the source input unscheduled, unless a timeout is passed which is honored (previously it was ignored and the syscall latency was used as the block time if it was over the thresholds). Adds a fallback in case a state is reached with no schedulable inputs yet some number of unscheduled inputs who would otherwise hang forever. Adds unit tests. Also tested on a large trace with many real-world cases of these direct switches and unschedulable threads. Issue: #6822
derekbruening
added a commit
that referenced
this issue
May 31, 2024
derekbruening
added a commit
that referenced
this issue
Jun 21, 2024
Adds support for threads starting out in an "unscheduled" state. This is accomplished by always reading ahead in each input and looking for a TRACE_MARKER_TYPE_SYSCALL_UNSCHEDULE marker *before* the first instruction. Normally such a marker indicates the invocation of a system call and is after the system call instruction; for start-unscheduled threads it is present at the system call exit at the start of the trace. Changes the scheduler's virtual method process_next_initial_record() to make the booleans on finding certain markers input-and-output parameters and moves filetype marker handling and timestamp recording into the function. This also fixes a problem where an input's initial next_timestamp was replaced with the 2nd timestamp if a subclass read ahead. The extra readahead causes complexities elsewhere which are addressed: + The reader caches the last cpuid to use for synthetic recores on skipping. + Generalizes the existing scheduler handling of readahead (the "recorded_in_schedule" field in input_info_t) to store a count of pre-read instructions, which will generally be either 0 or 1. Adds a new internal interface get_instr_ordinal() to get the input reader's instruction ordinal minus the pre-read count. Changes raw2trace's virtual function process_marker_additionally() to process_marker() and moves all marker processing (including timestamps, which are not markers in the raw format) there, to better support subclasses inserting start-unscheduled markers and deciding whether to insert new markers either before or after pre-existing markers. Adds a scheduler test for the new feature. Issue: #6822
derekbruening
added a commit
that referenced
this issue
Jun 22, 2024
Adds support for threads starting out in an "unscheduled" state. This is accomplished by always reading ahead in each input and looking for a TRACE_MARKER_TYPE_SYSCALL_UNSCHEDULE marker *before* the first instruction. Normally such a marker indicates the invocation of a system call and is after the system call instruction; for start-unscheduled threads it is present at the system call exit at the start of the trace. Changes the scheduler's virtual method process_next_initial_record() to make the booleans on finding certain markers input-and-output parameters and moves filetype marker handling and timestamp recording into the function. This also fixes a problem where an input's initial next_timestamp was replaced with the 2nd timestamp if a subclass read ahead. The extra readahead causes complexities elsewhere which are addressed: + The reader caches the last cpuid to use for synthetic recores on skipping. + Generalizes the existing scheduler handling of readahead (the "recorded_in_schedule" field in input_info_t) to store a count of pre-read instructions, which will generally be either 0 or 1. Adds a new internal interface get_instr_ordinal() to get the input reader's instruction ordinal minus the pre-read count. Changes raw2trace's virtual function process_marker_additionally() to process_marker() and moves all marker processing (including timestamps, which are not markers in the raw format) there, to better support subclasses inserting start-unscheduled markers and deciding whether to insert new markers either before or after pre-existing markers. Adds a scheduler test for the new feature. Issue: #6822
derekbruening
added a commit
that referenced
this issue
Jun 25, 2024
Adds a new virtual raw2trace per-record action point observe_entry_output() for use in subclasses to compute statistics or possibly trim nearly-empty start-unscheduled threads. Issue: #6822
derekbruening
added a commit
that referenced
this issue
Jun 25, 2024
Adds a new virtual raw2trace per-record action point observe_entry_output() for use in subclasses to compute statistics or possibly trim nearly-empty start-unscheduled threads. Issue: #6822
derekbruening
added a commit
that referenced
this issue
Jul 15, 2024
Adds a missing closing '>' to the drmemtrace view tool's output for a syscall that makes a thread become unscheduled. Issue: #6822
derekbruening
added a commit
that referenced
this issue
Jul 16, 2024
Adds a missing closing '>' to the drmemtrace view tool's output for a syscall that makes a thread become unscheduled. Issue: #6822
derekbruening
added a commit
that referenced
this issue
Sep 26, 2024
Fixes a bug for going-unscheduled actions where scaling a small timeout rounds it down to 0, which is then interpreted as an infinite timeout. Adds a unit test that fails without the fix. Issue: #6822
derekbruening
added a commit
that referenced
this issue
Sep 27, 2024
Fixes a bug for going-unscheduled actions where scaling a small timeout rounds it down to 0, which is then interpreted as an infinite timeout. Adds a unit test that fails without the fix. Issue: #6822
derekbruening
added a commit
that referenced
this issue
Sep 28, 2024
derekbruening
added a commit
that referenced
this issue
Oct 1, 2024
Adds a new drmemtrace scheduler option scheduler_options_t.honor_infinite_timeouts and a CLI parameter -sched_infinite_timeouts, both off by default. If turned on, these match the previous behavior. Changes the default behavior to use the (scaled per the scale parameter) maximum block timeout for indefinitely-unscheduled cases, rather than using an infinite timeout. This avoids waiting a long time for things like background threads that do nothing but wait the entire duration of a trace. Adds unit test variants for both infinite and max-timeout. Tested on a large application where this did not noticeably decrease the successful number of direct switches, but did reduce the idle time which was too high previously. Issue: #6822
derekbruening
added a commit
that referenced
this issue
Oct 2, 2024
Adds a new drmemtrace scheduler option scheduler_options_t.honor_infinite_timeouts and a CLI parameter -sched_infinite_timeouts, both off by default. If turned on, these match the previous behavior. Changes the default behavior to use the (scaled per the scale parameter) maximum block timeout for indefinitely-unscheduled cases, rather than using an infinite timeout. This avoids waiting a long time for things like background threads that do nothing but wait the entire duration of a trace. Adds unit test variants for both infinite and max-timeout. Tested on a large application where this did not noticeably decrease the successful number of direct switches, but did reduce the idle time which was too high previously. Issue: #6822
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
The drmemtrace scheduler currently has support for TRACE_MARKER_TYPE_DIRECT_SWITCH but it does not fully model "unscheduled" threads which will not run until explicitly woken up by another thread. This issue covers adding this support through new markers for waiting, resuming, and timeouts.
The text was updated successfully, but these errors were encountered: