Replace container of HPACK dynamic table from std::vector to std::deque

[masaori335]

The std::deque is an appropriate container for HPACK dynamic table. The use cases are below.
All of them are O(1) with std::deque.

• random access
• insertion at its beginning
• deletion at its end

[maskit]

Use of deque sounds reasonable to me. If I remember correctly the reason we didn't use it is that our data container policy only allowed use of std::vector from std data containers. If we allow it now, there's no reason to not use it.

It should be faster than current but did you check the performance?

[vmamidi]

What is the data container policy ?

…

On Oct 29, 2019, at 9:25 PM, Masakazu Kitajo ***@***.***> wrote:  Use of deque sounds reasonable to me. If I remember correctly the reason we didn't use it is that our data container policy only allowed use of std::vector from std data containers. If we allow it now, there's no reason to not use it. It should be faster than current but did you check the performance? — You are receiving this because your review was requested. Reply to this email directly, view it on GitHub, or unsubscribe.

[maskit]

What is the data container policy ?

Undocumented (or hard to find) policy. I'm not sure if there is a list of containers we can use, but the basic rule is this. https://issues.apache.org/jira/browse/TS-3770
As far as I know, vector and unordered_map are the only STL data containers we can use.

[masaori335]

As far as I know, ATS doesn't have "double-ended queue" nor any data structures which satisfy our needs here. IMO, we should add std::deque on the white list if it doesn't have. Also, we should write the policy on the cwiki or somewhere.

As for the benchmark, the throughput is almost same. However, from the dtrace, memmove on the HpackDynamicTable::_evict_overflowed_entries() like below is disappeared.

              libsystem_platform.dylib`_platform_memmove$VARIANT$Haswell+0x110
              traffic_server`HpackDynamicTable::_evict_overflowed_entries()+0xc9
              traffic_server`HpackDynamicTable::add_header_field(MIMEField const*)+0x6e

[vmamidi]

I don’t think you can see performance improvement with a regular throughput test. You have to do a test that involves lot of evictions.

…

On Oct 30, 2019, at 12:20 AM, Masaori Koshiba ***@***.***> wrote:  As far as I know, ATS doesn't have "double-ended queue" nor any data structures which satisfy our needs here. IMO, we should add std::deque on the white list if it doesn't have. Also, we should write the policy on the cwiki or somewhere. As for the benchmark, the throughput is almost same. However, from the dtrace, memmove on the HpackDynamicTable::_evict_overflowed_entries() like below is disappeared. libsystem_platform.dylib`_platform_memmove$VARIANT$Haswell+0x110 traffic_server`HpackDynamicTable::_evict_overflowed_entries()+0xc9 traffic_server`HpackDynamicTable::add_header_field(MIMEField const*)+0x6e — You are receiving this because your review was requested. Reply to this email directly, view it on GitHub, or unsubscribe.

[zwoop]

Cherry-picked to v9.0.x branch.