Reduce number of transmitted ObjectService.Search response messages

[cthulhu-rider]

Is your feature request related to a problem? Please describe.

object SEARCH operations are executed additively on all container nodes. Currently, server writes one response message to the stream per response from each individual node. When each node stores a few objects, this leads to a decrease in performance due to the overhead of additional messages when transmitting a fixed payload (list of object IDs for SEARCH)

Describe the solution you'd like

buffer the results up to the limit volume, and then flush them into the stream. Seems like this component is a good place to do

neofs-node/pkg/services/object/transport_splitter.go

Lines 151 to 184 in 9f4bf88

	func (s *searchStreamMsgSizeCtrl) Send(resp *object.SearchResponse) error {
	body := resp.GetBody()
	ids := body.GetIDList()
	var newResp *object.SearchResponse
	for ln := uint64(len(ids)); ; {
	if newResp == nil {
	newResp = new(object.SearchResponse)
	newResp.SetBody(body)
	}
	cut := s.addrAmount
	if cut > ln {
	cut = ln
	}
	body.SetIDList(ids[:cut])
	newResp.SetMetaHeader(resp.GetMetaHeader())
	newResp.SetVerificationHeader(resp.GetVerificationHeader())
	if err := s.stream.Send(newResp); err != nil {
	return err
	}
	ids = ids[cut:]
	if len(ids) == 0 {
	break
	}
	}
	return nil
	}

Describe alternatives you've considered

leave as it is. One of the possible benefits of the current approach with client POV is faster receipt of primary results from faster nodes. However, if the client understands the internal structure of the cluster so deeply and wants to more flexibly control operations on it, then either:

• it can manage cluster nodes itself - for this all information is available (netmap, containers)
• the protocol can be extended with a buffering bypass service flag upon special feature request (not yet)

Additional context

currently it's hard to detect specifics of such internal container queries. I've noticed current case during deep debug within #2722

[roman-khimov]

This requires both size and time constraints to balance latencies/bandwidth.