Performance of my python program goes down while using multiprocessing

I'm trying to make the following program run faster, so I used multiprocessing as suggested in the responses to the last question I asked, also on the related question. So, I've got the following program which runs very slowly than before:

import threading from threading import Thread from multiprocessing import Process def sim_QxD_word(query, document, model, alpha, outOfVocab, lock): #word_level sim_w = {} for q in set(query.split()): sim_w[q] = {} qE = [] if q in model.vocab: qE = model[q] elif q in outOfVocab: qE = outOfVocab[q] else: qE = numpy.random.rand(model.layer1_size) # random vector lock.acquire() outOfVocab[q] = qE lock.release() for d in set(document.split()): dE = [] if d in model.vocab: dE = model[d] elif d in outOfVocab: dE = outOfVocab[d] else: dE = numpy.random.rand(model.layer1_size) # random vector lock.acquire() outOfVocab[d] = dE lock.release() sim_w[q][d] = sim(qE,dE,alpha) return (sim_w, outOfVocab) def sim_QxD_sequences(query, document, model, outOfVocab, alpha, lock): #sequence_level # 1. extract document sequences document_sequences = [] for i in range(len(document.split())-len(query.split())): document_sequences.append(" ".join(document.split()[i:i+len(query.split())])) # 2. compute similarities with a query sentence lock.acquire() query_vec, outOfVocab = avg_sequenceToVec(query, model, outOfVocab, lock) lock.release() sim_QxD = {} for s in document_sequences: lock.acquire() s_vec, outOfVocab = avg_sequenceToVec(s, model, outOfVocab, lock) lock.release() sim_QxD[s] = sim(query_vec, s_vec, alpha) return (sim_QxD, outOfVocab) def word_level(q_clean, d_text, model, alpha, outOfVocab, out_w, q, ext_id, lock): print("in word_level") sim_w, outOfVocab = sim_QxD_word(q_clean, d_text, model, alpha, outOfVocab, lock) numpy.save(join(out_w, str(q)+ext_id+"word_interactions.npy"), sim_w) def sequence_level(q_clean, d_text, model, outOfVocab, alpha, out_s, q, ext_id, lock): print("in sequence_level") sim_s, outOfVocab = sim_QxD_sequences(q_clean, d_text, model, outOfVocab, alpha, lock) numpy.save(join(out_s, str(q)+ext_id+"sequence_interactions.npy"), sim_s) def extract_AllFeatures_parall(q_clean, d_text, model, alpha, outOfVocab, out_w, q, ext_id, out_s, lock): print("in extract_AllFeatures") thW=Process(target = word_level, args=(q_clean, d_text, model, alpha, outOfVocab, out_w, q, ext_id, lock,)) thS=Process(target = sequence_level, args=(q_clean, d_text, model, outOfVocab, alpha, out_s, q, ext_id, lock,)) thW.start() thS.start() thW.join() thS.join() def process_documents(documents, index, model, alpha, outOfVocab, out_w, out_s, queries, stemming, stoplist, q): print("in process_documents") q_clean = clean(queries[q],stemming, stoplist) lock = threading.Lock() for d in documents: ext_id, d_text = reaDoc(d, index) extract_AllFeatures_parall(q_clean, d_text, model, alpha, outOfVocab, out_w, q, ext_id, out_s, lock) outOfVocab={} # shared variable over all threads queries = {"1":"first query", ...} # can contain 200 elements .... threadsList = [] for q in queries.keys(): thread = Process(target = process_documents, args=(documents, index, model, alpha, outOfVocab, out_w, out_s, queries, stemming, stoplist, q,)) thread.start() threadsList.append(thread): for th in threadsList: th.join() 

But it seems like all the processes are running on the same node or like just one is running. I don't have a great experience on multi-processing with python, maybe the problem is about the use of the shared lock, because after some seconds of running the program I get this screen with "htop" command:

After that, just one node is running and heer is what's printed on the consol:

in process_documents in process_documents in extract_AllFeatures in extract_AllFeatures in word_level in word_level in sequence_level in sequence_level 

which means that functions are running, but it's very slow.