[
https://issues.apache.org/jira/browse/IMPALA-1638?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Tim Armstrong resolved IMPALA-1638.
-----------------------------------
Resolution: Won't Do
I was able to get this working but it didn't have a noticeable impact on
performance as far as I could tell. I benchmarked impala-profile-tool on a
release build using this:
{noformat}
./buildall.sh -release -noclean -notests -skiptests -ninja && ninja
impala-profile-tool && while cat test-profile-logs/impala_profile_log* | time
./be/build/latest/util/impala-profile-tool --profile_format=text > /dev/
null; do date; done
{noformat}
Before:
{noformat}
4.65user 0.54system 0:05.22elapsed 99%CPU (0avgtext+0avgdata
60672maxresident)k
0inputs+0outputs (0major+83026minor)pagefaults 0swaps
Tue 15 Dec 2020 03:05:48 PM PST
4.53user 0.63system 0:05.19elapsed 99%CPU (0avgtext+0avgdata
60428maxresident)k
0inputs+0outputs (0major+83018minor)pagefaults 0swaps
{noformat}
With this change:
{noformat}
4.67user 0.47system 0:05.17elapsed 99%CPU (0avgtext+0avgdata
60560maxresident)k
0inputs+0outputs (0major+83015minor)pagefaults 0swaps
4.54user 0.52system 0:05.08elapsed 99%CPU (0avgtext+0avgdata
60548maxresident)k
0inputs+0outputs (0major+83010minor)pagefaults 0swaps
{noformat}
Diff:
{noformat}
tarmstrong@tarmstrong-Precision-7540:~/impala/impala$ git diff HEAD^^
diff --git a/be/src/rpc/thrift-util.h b/be/src/rpc/thrift-util.h
index 05e7e5544..086c690ed 100644
--- a/be/src/rpc/thrift-util.h
+++ b/be/src/rpc/thrift-util.h
@@ -21,6 +21,7 @@
#include <boost/shared_ptr.hpp>
#include <thrift/protocol/TBinaryProtocol.h>
+#include <thrift/protocol/TCompactProtocol.h>
#include <sstream>
#include <vector>
#include <thrift/TApplicationException.h>
@@ -112,7 +113,11 @@ Status DeserializeThriftMsg(const uint8_t* buf, uint32_t*
len, bool compact,
boost::shared_ptr<apache::thrift::protocol::TProtocol> tproto =
CreateDeserializeProtocol(tmem_transport, compact);
try {
- deserialized_msg->read(tproto.get());
+ if (compact) {
+
deserialized_msg->read(static_cast<apache::thrift::protocol::TCompactProtocol*>(tproto.get()));
+ } else {
+ deserialized_msg->read(tproto.get());
+ }
} catch (std::exception& e) {
std::stringstream msg;
msg << "couldn't deserialize thrift msg:\n" << e.what();
diff --git a/common/thrift/CMakeLists.txt b/common/thrift/CMakeLists.txt
index 9b14ab58d..ba282363e 100644
--- a/common/thrift/CMakeLists.txt
+++ b/common/thrift/CMakeLists.txt
@@ -59,7 +59,7 @@ function(THRIFT_GEN VAR)
# The java dependency is handled by maven.
# We need to generate C++ src file for the parent dependencies using the
"-r" option.
set(CPP_ARGS ${THRIFT_INCLUDE_DIR_OPTION}
- --gen cpp:moveable_types,no_default_operators -o ${BE_OUTPUT_DIR})
+ --gen cpp:moveable_types,no_default_operators,templates -o
${BE_OUTPUT_DIR})
IF (THRIFT_FILE STREQUAL "beeswax.thrift")
set(CPP_ARGS -r ${CPP_ARGS})
ENDIF(THRIFT_FILE STREQUAL "beeswax.thrift")
{noformat}
In perf top I could see that switch from indirect to direct calls but most of
the time seemed to be spent in memory allocation regardless.
> Investigate using c++ templates option when generating thrift and increasing
> the transport buffer
> -------------------------------------------------------------------------------------------------
>
> Key: IMPALA-1638
> URL: https://issues.apache.org/jira/browse/IMPALA-1638
> Project: IMPALA
> Issue Type: Task
> Components: Distributed Exec
> Affects Versions: Impala 2.2
> Reporter: casey
> Assignee: Tim Armstrong
> Priority: Minor
> Labels: performance
>
> While investigating the performance of "select * from tpch.lineitem" thrift
> seemed to be very slow. I did a benchmark comparing thrift and captnproto to
> transfer a total of 1gb using 1mb responses over the loopback. captnproto
> took ~0.5 seconds where thrift took ~6 seconds. Thrift was setup similar to
> how it's used in Impala. Todd was able to get the thrift timing down to ~1.7
> seconds with a few simple tweaks that aren't used by Impala. The improvements
> are
> 1) Generate code using templates. Without this, thrift generates inheritance
> style code which results in a virtual call to read and write every data point
> (such as an int).
> 2) Use a framed transport. The problem was that even when using the buffered
> transport, the default buffer was too small (Impala also uses the defualt
> buffer size). It's possible all we need to do is increase the buffer size.
> Testing should be done on a real cluster since the loopback could give very
> different results.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
