unofficial mirror of meta@public-inbox.org
 help / color / mirror / Atom feed
* [TESTING] WIP - parallel shards on HDD with sequential flush
@ 2020-08-12  9:34 Eric Wong
  2020-08-12 20:26 ` Eric Wong
  0 siblings, 1 reply; 2+ messages in thread
From: Eric Wong @ 2020-08-12  9:34 UTC (permalink / raw)
  To: meta

Frequent flushing to save RAM with HDD is horrible with random
writes Xapian tends to do; especially when parallelized

I think just making the Xapian commits in sequence while the
random reads + in-memory changes are still parallelized
is doable, though...

With this, --no-fsync may even be detrimental because Linux
defaults to excessive writeback buffering.

This could take some time to test (using ext4 on an HDD).  So
far it's showing decent CPU usage on a 4 core system and
sequential shard commits only show one process in D state at a
time and it seems to be going at a decent rate...

Waiting to see if it slows down as the Xapian DBs get bigger...
---
 lib/PublicInbox/V2Writable.pm | 34 +++++++++++++++++++++++++++-------
 1 file changed, 27 insertions(+), 7 deletions(-)

diff --git a/lib/PublicInbox/V2Writable.pm b/lib/PublicInbox/V2Writable.pm
index d99e476a..51985a58 100644
--- a/lib/PublicInbox/V2Writable.pm
+++ b/lib/PublicInbox/V2Writable.pm
@@ -20,6 +20,7 @@ use PublicInbox::Spawn qw(spawn popen_rd);
 use PublicInbox::SearchIdx qw(log2stack crlf_adjust is_ancestor check_size);
 use IO::Handle; # ->autoflush
 use File::Temp ();
+use Carp qw(croak);
 
 my $OID = qr/[a-f0-9]{40,}/;
 # an estimate of the post-packed size to the raw uncompressed size
@@ -114,6 +115,7 @@ sub new {
 		total_bytes => 0,
 		current_info => '',
 		xpfx => $xpfx,
+		seq_flush => 1,
 		over => PublicInbox::OverIdx->new("$xpfx/over.sqlite3"),
 		lock_path => "$dir/inbox.lock",
 		# limit each git repo (epoch) to 1GB or so
@@ -593,14 +595,20 @@ sub barrier_init {
 	my $barrier = { map { $_ => 1 } (0..$n) };
 }
 
+sub barrier_wait_1 ($) {
+	my ($r) = @_;
+	defined(my $line = readline($r)) or croak "EOF on barrier_wait: $!";
+	$line =~ /\Abarrier (\d+)/ or croak "bad line on barrier_wait: $line";
+	$1;
+}
+
 sub barrier_wait {
 	my ($self, $barrier) = @_;
 	my $bnote = $self->{bnote} or return;
 	my $r = $bnote->[0];
 	while (scalar keys %$barrier) {
-		defined(my $l = readline($r)) or die "EOF on barrier_wait: $!";
-		$l =~ /\Abarrier (\d+)/ or die "bad line on barrier_wait: $l";
-		delete $barrier->{$1} or die "bad shard[$1] on barrier wait";
+		my $i = barrier_wait_1($r);
+		delete($barrier->{$i}) or die "bad shard[$i] on barrier wait";
 	}
 }
 
@@ -609,7 +617,7 @@ sub checkpoint ($;$) {
 	my ($self, $wait) = @_;
 
 	if (my $im = $self->{im}) {
-		if ($wait) {
+		if ($wait || $self->{seq_flush}) {
 			$im->barrier;
 		} else {
 			$im->checkpoint;
@@ -626,15 +634,27 @@ sub checkpoint ($;$) {
 		$self->{over}->commit_lazy;
 
 		# Now deal with Xapian
-		if ($wait) {
+		if ($self->{seq_flush} && $self->{parallel}) {
+			my $r = $self->{bnote}->[0] or
+					die "BUG: {bnote} missing";
+			my $i = 0;
+			for (@$shards) {
+				$_->remote_barrier;
+				my $j = barrier_wait_1($r);
+				$i == $j or die <<EOF;
+BUG: bad shard[$j] on barrier wait (expected $i)
+EOF
+				$i++
+			}
+		} elsif ($wait) { # parallel flush w/ shared barrier
 			my $barrier = $self->barrier_init(scalar @$shards);
 
 			# each shard needs to issue a barrier command
 			$_->remote_barrier for @$shards;
 
 			# wait for each Xapian shard
-			$self->barrier_wait($barrier);
-		} else {
+			barrier_wait($self, $barrier);
+		} else { # sequential flush with 1 process
 			$_->remote_commit for @$shards;
 		}
 

^ permalink raw reply related	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2020-08-12 20:26 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-08-12  9:34 [TESTING] WIP - parallel shards on HDD with sequential flush Eric Wong
2020-08-12 20:26 ` Eric Wong

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).