The analytical algorithm performance doesn't meet the expectation

[vegetableysm]

I tested pagerank and sssp algorithm on GraphScope and Gemini. And found that the Gemini performed much faster than the GraphScope. Here is the result of pagerank test with the dataset of soc-LiveJournal1 . The iterations is 20:

platform	time usage
GraphScope	1.4s
Gemini	0.37s
Libgrape-lite	0.15s

The script of GraphScope

import graphscope
import time
import datetime

sess = graphscope.session(num_workers=1, cluster_type='hosts')

graph = sess.g()

start = time.time()
graph = graph.add_edges('../data_set/live_journal/soc-livejournal.csv', src_label='v', dst_label='v', properties = [])
end = time.time()

print("Loading time: %f" % (end - start))

start = time.time()
ret1 = graphscope.pagerank(graph, max_round = 20)
end = time.time()

print("Running time: %f" % (end - start))

print(ret1.to_dataframe(selector={'id': 'v.id', 'label': 'r'}))

sess.close()

The command of running libgrape-lite
mpirun -n 1 ./run_app --vfile ../../data_set/live_journal/soc-livejournal.vertex.csv --efile ../../data_set/live_journal/soc-livejournal.mtx --application pagerank --out_prefix ./output_pagerank --directed -pr_mr 20

The running command of Gemini
./toolkits/pagerank ./data_set/live_journal/soc-livejournal.binarye 4033137 20

Each of the above three tests sets one partition.

The problem is that the graphscope is 10 times worse than the libgrape-lite. I don't know if my test script is wrong, please advise. Thanks!

[welcome]

Thanks for opening your first issue here! Be sure to follow the issue template! And a maintainer will get back to you shortly!
Please feel free to contact us on DingTalk, WeChat account(graphscope) or Slack. We are happy to answer your questions responsively.

[siyuan0322]

One reason of that is it may includes the compilation time and projection time.
Other reasons are under investigation.

[vegetableysm]

The first ten lines of edge file:

src,dst
2,1
3,1
5,1
6,1
7,1
8,1
9,1
10,1
16,1

[siyuan0322]

Found that the timing method includes the python codes to assemble the op, and the round trip time of RPC, the dynamic loading of libraries, thus add much over head to querying time (which is less than 1 second in this experiment), so the overhead is huge.

Add a new log to print the actual evaluating time of application in the grape_engine, which should as the realiable metrics.