Details
Description
We've observed issues where the masters are slow to respond. Two perf traces collected while the masters were slow to respond:
25.84% [kernel] [k] default_send_IPI_mask_sequence_phys 20.44% [kernel] [k] native_write_msr_safe 4.54% [kernel] [k] _raw_spin_lock 2.95% libc-2.5.so [.] _int_malloc 1.82% libc-2.5.so [.] malloc 1.55% [kernel] [k] apic_timer_interrupt 1.36% libc-2.5.so [.] _int_free
29.03% [kernel] [k] default_send_IPI_mask_sequence_phys 9.64% [kernel] [k] _raw_spin_lock 7.38% [kernel] [k] native_write_msr_safe 2.43% libc-2.5.so [.] _int_malloc 2.05% libc-2.5.so [.] _int_free 1.67% [kernel] [k] apic_timer_interrupt 1.58% libc-2.5.so [.] malloc
These have been found to be attributed to the posix_fadvise calls made by glog. We can disable these via the environment:
GLOG_DEFINE_bool(drop_log_memory, true, "Drop in-memory buffers of log contents. " "Logs can grow very quickly and they are rarely read before they " "need to be evicted from memory. Instead, drop them from memory " "as soon as they are flushed to disk.");
if (FLAGS_drop_log_memory) { if (file_length_ >= logging::kPageSize) { // don't evict the most recent page uint32 len = file_length_ & ~(logging::kPageSize - 1); posix_fadvise(fileno(file_), 0, len, POSIX_FADV_DONTNEED); } }
We should set GLOG_drop_log_memory=false prior to making our call to google::InitGoogleLogging, to avoid others running into this issue.
Attachments
Issue Links
- is related to
-
IMPALA-7252 Backport rate limiting of fadvise calls into toolchain glog
- Resolved
-
MESOS-1375 Log rotation capable
- Accepted
-
MESOS-9680 Remove automatic disablement of GLOG_drop_log_memory.
- Accepted